Enhanced mask-region-based segmentation with parallel CNN for object detection in traffic scenarios

1 Introduction

As autonomous driving technology evolves rapidly, road-object detection has become increasingly important in research. Detecting objects within road environments is considered a vital component of the environmental perception system in autonomous vehicles (Jiang et al., 2023). For devices with limited processing capabilities, it is essential to detect objects swiftly and accurately in real-world traffic conditions to ensure safe and reliable driving decisions (Liang et al., 2022). Deep-learning-based approaches are increasingly replacing traditional algorithms for object detection due to their enhanced precision and generalization ability (Guo et al., 2020). These methods have shown promising results in identifying objects in traffic environments. However, significant challenges persist (Zhang et al., 2022). In the current landscape of autonomous driving perception tasks, factors such as varying weather conditions, changing lighting, object occlusions, and the presence of small objects in real traffic scenarios introduce considerable uncertainty, which, in turn, affects the accuracy of road object detection systems (Li A. et al., 2023).

In traffic scenarios, there are numerous small objects at a distance which present challenges for detection due to the limited feature information arising from their small size (Lai et al., 2023). This makes them more difficult to accurately detect. Research in small object detection has led to the development of several approaches, including the deconvolutional single-shot detector (DSSD), scale normalization for image pyramids (SNIP), and high-resolution detection network (HRDNet) (Ye et al., 2020). The DSSD algorithm improves small object detection performance by employing an enhanced feature extraction network and incorporating contextual information (He et al., 2022). The SNIP algorithm introduces an innovative training method—scale normalization for image pyramids—which selectively back-propagates gradients based on the image scale of objects of different sizes, thus improving the detection of smaller objects (Patil et al., 2024). The HRDNet algorithm uses high-resolution inputs in a shallow network to retain more positional information while feeding low-resolution inputs into a deeper network for better semantic extraction (Liu et al., 2024). By processing features at various resolutions, HRDNet increases small object detection while also preserving efficacy for large and medium objects. Each of these algorithms offers distinct advantages and has its own set of limitations (Li S. et al., 2023).

Various methods have been proposed to tackle the issue of low detection accuracy in weakly supervised learning (Wu et al., 2020). One common approach involves using online instance classifier refinement (OICR) to iteratively refine bounding boxes, followed by further training of the fully supervised detector with the generated pseudo ground truth (Rezaei et al., 2023). However, this method still tends to focus on discriminative parts, leading to some limitations. Another notable approach, W2F, utilizes pseudo boxes (generated by existing weakly supervised detectors) as ground truth, effectively converting the weakly supervised detection problem into a supervised one (Gählert et al., 2020). However, this approach is still limited by the inaccuracies in the pseudo ground truth. Hence, this study proposes a novel normalized parallel convolutional neural network for object detection in traffic scenarios (NPC-ODT). The main contributions are as follows:

• contributing an MBA-MRCNN model for segmentation that adopts a bottleneck attention mechanism for better handling of complex images with varying object size;
• extracting a UC-HoS-based feature that enhances overall clarity of the image by eliminating imperfections before processing;
• proposing the NPCNN model that adopts normalized pTanh response batch normalization to ensure stable output by enhancing generalization and efficiency.

The remainder of this article is organized as follows. Section 2 reviews existing research related to object detection in traffic scenarios. Section 3 presents the proposed method. Section 4 examines the developed method and compares it with existing research, and Section 5 presents the conclusion.

2 Literature review

The literature on conventional studies related to object detection in traffic scenarios using different approaches and their experimental outcomes is now discussed.

Qiu et al. (2023) proposed the IDOD-YOLOV7 framework for object detection in foggy traffic environments. It combined an enhancement module (SAIP) and an image defogging module (AOD), jointly optimized with YOLOV7. The system improved low-light foggy images by enhancing image quality and performed object detection through end-to-end learning. A low-light FTOD was created to address domain transfer issues.

Mohamed and Chuan (2022) proposed a PAE model for single-stage object detection, which detected and extracted appearance embeddings using a shared model. They also presented a training structure that works with a unique image and no re-recognition annotations, reducing labeling effort and boosting real-world applicability. It was tested on various datasets and showed comparable outcomes to SOTA models.

Li et al. (2021) introduced an end-to-end DL framework with a driving-behavior-aware model that leveraged spatial-temporal details from consecutive frames. The model learned driving performance by analyzing deviations in 2D center points, rotation, translation, and depth in parallel. In contrast to earlier methods, the approach models driving behavior through reasoning about the possible motion trajectories of the board.

Niu et al. (2024) introduced DRUformer, a transformer-based multi-modal model for important object detection in driving scenarios. It considered the relationships between participants and incorporated driving intention. Furthermore, Niu et al. (2024) evaluated its performance on the DRAMA dataset, comparing it to other SOTA models, and performed a qualitative analysis showing its effectiveness across various road scenarios and object classes.

Zhao et al. (2024) introduced Z-YOLOv8s, which aims to improve both accuracy and real-time performance under real-world conditions. The model enhanced spatial feature extraction by integrating the RepViT, C2f, and LSKNet into the YOLOv8s framework. Furthermore, the YOLOv8s backbone was enhanced with SPD-Conv to expand small object recognition.

Garcia-Aguilar et al. (2023) proposed an automatic method for detecting small-scale objects in traffic sequences. Initially, vehicle patterns were generated from frames by pre-trained and super-resolution detection networks. The model was then reinstructed with this data to adapt to the scene, and was then applied in real-time to the video stream.

Lian et al. (2021) proposed a small object detection method for traffic scenarios using attention feature fusion. They designed a MS-CAB to aggregate information at global and local scales and introduced an AFFB to integrate contextual information across layers. Replacing the linear fusion scheme with the AFFB improved detection accuracy, achieving higher mAP than YOLOv5s while maintaining real-time performance.

He et al. (2021) introduced SA-MobileNeXt, a lightweight backbone with channel and spatial attention. This approach enhanced feature extraction with the SCSA module in the SGBlock, adding minimal parameters. Moreover, they used Mosaic-Mixup for data augmentation and proposed the MFEF network for better feature fusion. The SiLU activation optimized the CBL and SGBlock modules, speeding up convergence.

Table 1 focuses on characteristics and challenges in existing research related to object detection in traffic scenarios.

TABLE 1. Review of existing studies.

2.1 Problem statement

The objective of this project was to develop a robust object detection system for identifying and localizing various objects in traffic scenarios, such as vehicles, traffic signs, pedestrians, cyclists, and other relevant objects. The challenge lay in creating an efficient, real-time detection model that can operate reliably in urban, suburban, and rural settings to support various. For example, the integration of components such as SAIP and AOD in some models increases both time and spatial complexity (Qiu et al., 2023), while many methods fail to effectively balance accuracy, speed, and model complexity, as observed in existing object detection models (He et al., 2021). Additionally, many methods still struggle with handling motion prediction and dynamic tracking, which are crucial for real-time traffic safety, as noted by the potential benefits of using GANs to mimic object motion for a more effective training approach (Mohamed and Chuan, 2022). The system should be capable of high accuracy with minimal latency to ensure timely responses and provide essential information for subsequent tasks such as tracking, behavior prediction, and decision-making in intelligent transportation systems.

3 Proposed methodology

Object detection in traffic scenarios is a vital component of computer vision, essential for intelligent transportation, autonomous driving, and traffic control systems. It focuses on identifying and locating objects such as vehicles, pedestrians, traffic signs, bicycles, and road infrastructure in images or video captured by vehicle-mounted or roadside cameras. The task is challenging due to factors such as adverse weather, variable lighting, object occlusions, and high scene dynamics. Overcoming these challenges is key to building reliable autonomous systems capable of safely navigating complex real-world environments. This study proposes a novel normalized parallel convolutional neural network for object detection in traffic scenarios (NPC-ODT).

As shown in Figure 1, the process of object detection in traffic scenarios begins with an input video V i n p, which is first divided into individual frames I i f = I 1 f , I 2 f , . . . , I n f for detailed analysis. Each frame undergoes preprocessing using the Wiener filter, which helps reduce noise while maintaining important image details, thereby enhancing the quality of input data. Following this, segmentation is conducted using a modified bottleneck attention-based mask recurrent convolutional neural network (MBA-MRCNN), which effectively isolates regions of interest such as vehicles, pedestrians, and road signs from the background. Once the relevant segments are identified, feature extraction is carried out based on a hierarchy of methods. This includes updated center pixel in hierarchy of skeleton (UC-HoS) that captures shape-related cues, color-based features for object differentiation, and motion estimation to track moving elements across frames. These extracted features are then fed into NPCNN, which is trained to recognize and classify objects within the scene. The final output consists of accurately detected and labeled objects in each frame, enabling further tasks such as tracking or behavior analysis.

3.1 Preprocessing

The preprocessing step in object detection is crucial for refining the quality of the raw input data (typically video frames I i f) before it is fed into any segmentation or detection models. This stage aims to clean the data, remove noise, and enhance relevant features to facilitate more accurate analysis. One of the main tasks in preprocessing is reducing noise in the image, which can interfere with the subsequent detection and analysis stages. In traffic scenarios, noise can arise from various sources, such as camera sensors, weather conditions, or motion blur. The Wiener filter is commonly used for this purpose. It works by estimating the local noise characteristics and adjusting the pixel values to reduce noise while preserving important details (such as edges and textures) in the image.

The Wiener filter (Kalaivani and Asnath Victy Phamila, 2018) is an additive noise removal technique that utilizes a low-pass, pixel-wise adaptive method based on local neighborhood information. The method estimates the power spectra of both the noisy and original images by applying Fourier transforms to their autocorrelation functions. The filter then applies a mask to each noisy pixel, calculating the local mean (μ) and variance (σ²) for each pixel. Leveraging these values along with the noise power, the filter iteratively computes updated pixel values for every location in the noisy image. This method effectively balances noise smoothing and deconvolution through inverse filtering, making it an optimal solution for noise reduction. The performance of the Wiener filter is typically evaluated by comparing the PSNR and correlation coefficient values before and after filtering. Thus, the preprocessed image is indicated as p r e I f.

3.2 Segmentation

The segmentation step in object detection plays a critical role in isolating the relevant objects from the background in each frame. Segmentation divides an image p r e I f into multiple parts or “segments” that are more manageable and meaningful for further investigation. The main goal of segmentation is to separate objects of interest (such as cars, people, and road signs) from the background (such as the road, trees, and sky) to enable more accurate detection and recognition in later steps. Effective segmentation helps in focusing only on the key elements in a scene, reducing the complexity for feature extraction and object detection.

This study utilizes a modified bottleneck attention-based mask recurrent convolutional neural network (MBA-MRCNN) for segmentation. The MBA-MRCNN is the extension of Mask R-CNN model (Aabha et al., 2022), which is designed for object detection and segmentation. It starts with an input image p r e I f, which is managed by a CNN to extract feature maps. A region proposal network then identifies potential RoIs from these feature maps, which are passed through RoI Align for precise spatial alignment. The aligned regions are further processed by FC layers, followed by two distinct branches: one for object classification using a softmax layer and another for bounding box regression to fine-tune the location of detected objects. Additionally, the model has a dedicated branch for generating masks for each object through convolutional layers, enabling instance segmentation. Overall, this architecture is optimized to perform both object detection and segmentation, providing class labels and pixel-level masks for each identified object. However, it treats all regions equally, leading to inefficient feature extraction. It requires more training time due to less efficient feature handling. Due to the lack of focus on this important area, this model struggles with complex or crowded scenarios, which may suffer from lower accuracy in challenging segmentation tasks.

To address these issues, an MBA-MRCNN model is introduced that reduces training time by focusing only on relevant features. It is more adaptable to distinct scenarios with varying object size and complexities and achieves potentially maximal accuracy due to selected attention to important regions. The structural representation of MBA-MRCNN model is provided in Figure 2.

The process of MBA-MRCNN begins with a preprocessed image p r e I f being passed through a ResNet-50 model to extract feature maps. These feature maps are then refined using a MBA module to enhance the features further. The process of the MBA module begins with an input tensor T f, which is first passed through a global average pooling layer to condense spatial information into a compact form. This pooled output is then passed through a FC layer, followed by another FC layer with the number of channels reduced by a factor of “r.” A channel attention mechanism is applied, which focuses on the important channels by assigning weights to each channel based on its significance. Simultaneously, a second path processes the tensor through a 1×1 convolution, followed by a 3×3 convolution and another 1×1 convolution, which is designed to capture spatial attention by focusing on important spatial regions within the feature map. These outputs from both the spatial and channel attention are then added together. This combined output is then fed into a bottleneck attention to further refine the attention mechanism. Finally, the output of the bottleneck attention module is multiplied by the original input tensor, and the result is added back to the input tensor, producing the final output f M B A.

The feature map function f M B A ∈ G C × H × W, obtained from the depthwise separable convolution layer, is used by BAM to generate a 3D attention map A T f ∈ G C × H × W. Subsequently, the modified feature map f M B A is derived according to Equation 1. Here, ⊗ denotes element-wise multiplication.

f M B A = T f + T f ⊗ A T f . (1)

The BAM architecture integrates the attention mechanism with the learning block to enhance gradient flow. It starts by calculating the channel attention function A c T f ∈ G C and the spatial attention function to create a powerful and efficient module. Additionally, a sigmoid function A s T f ∈ G H × W is applied. As detailed in Equation 2, the attention map A T f is then derived by summing both the channel and spatial attention values.where σ * denotes modified sigma calculation, A c T f refers to channel attention function ∈ G C, and A s T f denotes spatial attention function ∈ G H × W. Conventional sigmoid (Ocal, 2025) has a vanishing gradient problem. To overcome this, a modified sigma is introduced as defined in Equation 3.where S i g m o i d Im x = ϑ 1 + e − ϑ x + δ, s w i s h x = x 1 + e − x, ϑ = 0.5, δ = 1, and χ = 7.

A T f = σ * A c T f + A s T f , (2)

σ * = ϑ x − χ + S i g m o i d Im χ , x ≥ χ S i g m o i d Im x , − χ < x < χ ϑ x + χ + s w i s h x , x ≤ χ , (3)

This flow emphasizes selective attention at both the channel and spatial levels, allowing the model to focus on the most relevant features and given to RPN. The output from a RPN is then applied to generate potential RoIs from the feature maps. The outputs from both the RPN and the feature maps are fed into the RoI Align layer for precise spatial alignment of the regions. Following this, the network branches into two distinct paths. The first focuses on object classification, which involves a sequence of a convolutional layer, batch normalization (BN), and Leaky ReLU activation, followed by two fully connected (FC) layers (FC1 and FC2), to predict the class of the object, and a softmax layer, culminating in a bounding box regression to refine the object’s location. The second path is dedicated to instance segmentation and includes a sequence of convolutional layers, BN, ReLU, and Leaky ReLU activations, and ends with a final convolutional layer that outputs a segmentation mask for each detected object. This architecture enables the model to simultaneously perform object detection by classifying and localizing objects and instance segmentation by producing pixel-level masks for each object. Thereby, the segmented image is indicated by I f s e g. In contrast to the classical Mask R-CNN algorithm, the MBA-MRCNN model uses the modified bottleneck attention block that implements both spatial and channel attention via the introduced modified sigmoid function. Thus, the MBA-MRCNN can recalibrate features adaptively prior to proposing regions, which improves segmentation in the traffic scene.

3.3 Feature extraction

The feature extraction step is a crucial part of the object detection pipeline, where specific characteristics of the segmented objects I f s e g are extracted and used to distinguish between different objects and identify important patterns in the scene. In traffic scenarios, feature extraction typically involves multiple levels of analysis, each focusing on different aspects of the objects and their behaviors. The primary goal of the feature extraction step is to identify distinct characteristics or patterns that can help classify and track the objects of interest in the traffic scenarios.

3.3.1 Updated center pixel in hierarchy of skeleton (UC-HoS)

“Hierarchy-of-skeleton-based feature extraction” (Chen et al., 2012) refers to a structured approach to extracting features from the skeletal representation of an object or human body in the segmented image. HoS uses a parallel thinning algorithm. The process involves representing an object or human body as a set of interconnected joints (skeleton) and then extracting hierarchical features at various levels of granularity to improve detection and recognition accuracy.

However, the skeletonized image can have artifacts, inaccuracies, or weaker boundaries due to noise and edge issues. This can struggle with complex images, leading to inaccurate or incomplete skeletons. To overcome these challenges, UC-HoS is introduced to enhance edge definition by focusing on the center pixel, resulting in clearer and sharper boundaries. This leads to a cleaner skeletonization process with more robust and accurate outcomes. We update the center pixel of the segmented image with the following procedures.

Consider a small 3×3 grid of pixel values representing a tiny portion of an image (Figure 3). The center pixel is updated by computing the weighted sum of the pixel at the center and its four adjacent pixels.

Assuming the center pixel is I f s e g u , v, the distribution of weights between the center pixel and its four neighboring pixels is as provided in Equations 4–8:

ϖ u , v = 0.4 , (4)

ϖ u + 1 , v = 0.15 , (5)

ϖ x − 1 , y = 0.15 , (6)

ϖ u , v + 1 = 0.15 , (7)

ϖ u , v − 1 = 0.15 . (8)

Each pixel is assigned a weight within the range 0–1, ensuring that the total weighted sum of the center pixel and its four adjacent pixels adds up to 1. The weighted sum for the pixel located at x , y is determined using the following Equation 9:

I ϖ f s e g u , v = I f u , v ϖ u , v + s e g I f u + 1 , v . s e g ϖ u + 1 , v + I f s e g u − 1 , v . ϖ u − 1 , v + I f s e g u , v + 1 . ϖ u , v + 1 + I f u , v − 1 . ϖ s e g u , v − 1 (9)

The following Equation 10 is used to update the value of center pixel:

I ϖ f s e g u , v = I f u , v , s e g I ϖ f u , v , s e g I ϖ f u , v − I f u , v s e g ≤ T h r s e g I ϖ f u , v − I f u , v s e g > T h r s e g (10)

Instead of setting fixed threshold (i.e., T h r = 6) (Udupa, 2023), this study adopts a small constant value, which is multiplied with the standard deviation of the pixel value within this neighborhood as defined in Equation 11:where K is the small constant value of 0.1 and s t d denotes standard deviation of the pixel value within this neighborhood.

T h r = K * s t d (11)

Based on this, the boundary points need to be processed using the computed new center pixel value ρ 1 (Figure 3), in accordance with the subsequent procedures.

For a boundary point and its 8-connected neighborhood, the center point is identified as ρ 1, with the surrounding points labeled ρ 2 to ρ 9 in a clockwise direction. The first step involves selecting all points that satisfy the conditions outlined below in Equations 12–15. where n ρ 1 denotes non-zero adjacent pixel points count and S ρ 1 indicates changes in series count. Once all boundary points had been examined, all the marked points were removed.

C n d t n 1 : 2 ≤ n ρ 1 ≤ 6 (12)

C n d t n 2 : S ρ 1 = 1 (13)

C n d t n 3 : ρ 2 × ρ 4 × ρ 6 = 0 (14)

C n d t n 4 : ρ 4 × ρ 6 × ρ 8 = 0 (15)

As in step (i), the condition C n d t n 3 was redefined as ρ 2 × ρ 4 × ρ 6 = 0 and condition C n d t n 4 as ρ 4 × ρ 6 × ρ 8 = 0. The identified points were discarded following the inspection.

The two steps above constitute one iteration. This process continues until no points meet the specified requirements, and the remaining points form the regional skeleton as S k e f. The proposed UC-HoS introduces adaptive center-pixel updating before skeletonization, which preserves local structural continuity and generates cleaner skeletal representations.

3.3.2 Color features

Color features (Zenggang et al., 2021) C f play a crucial role in object detection within traffic scenarios, enabling the differentiation of objects such as vehicles, pedestrians, traffic lights, and road signs from their surroundings and from one another. The HSV color space is particularly effective as it isolates hue (color) from intensity (brightness), improving the system’s robustness against changing environmental conditions. Moreover, using color histograms and multi-channel information enhances detection precision by capturing the prominent color characteristics of objects in busy, dynamic traffic situations.

3.3.3 Motion estimation-based feature

Motion estimation-based feature extraction involves identifying movement patterns between consecutive video frames by comparing small regions called blocks. The block matching (BM) algorithm (Betka et al., 2019) is a fundamental technique in motion estimation used in video processing. The method employs the full search algorithm (FSA), which involves dividing both the current and reference video frames into non-overlapping square blocks. Each block in the current frame is matched to the most similar block within a designated window in the adjacent reference frame. The similarity between blocks is typically measured using cost functions such as the SAD. The algorithm iteratively compares each block in the current frame with all candidate blocks in the reference frame, computing the cost for each and selecting the one with the minimum value as the best match. The difference in position between the current block and its best-matched block defines the motion vector. This process is repeated for every block in every frame, generating a set of motion vectors that represent object movement across frames. Although highly accurate, the BM algorithm can be computationally intensive due to the exhaustive search required, especially with larger search windows. Thus, the motion-estimated-based feature can be indicated as M E f. Then, the entire feature extracted from s e g I f is determined as E f = S k e f , C f , M E f.

3.4 Object detection via NPCNN

The object detection step is the final stage in the object detection pipeline, where its objective is to identify and locate specific objects within an image or video frame. In this step, the normalized parallel convolutional neural network (NPCNN) uses the extracted features E f to classify objects. The NPCNN is the variant of PCNN (Ye et al., 2023) processing an image through multiple parallel convolutional branches, each focusing on different aspects such as 1D, 2D, or dilated convolutions. The outputs of these branches are fused in a fusion layer, combining the features into a unified representation. This fused output then passes through FC layers for further refinement, followed by a dropout layer to avoid overfitting. Finally, the network produces the output through a Softmax layer for classification or a linear output for regression tasks. This architecture enables efficient feature extraction from diverse perspectives. The main limitation of PCNN is that it learns the redundant features due to different branches. This can reduce the overall effectiveness of the model. The texture appearance of both branches does not have distinct features to differentiate them. To address this limitation, NPCNN is introduced and their diagrammatic representation is provided in Figure 4.

The process begins with the input feature, which is passed through two separate convolutional blocks. The first block is a 1D CNN block, where the input undergoes a sequence of operations: convolution 1, followed by normalized pTanh response batch normalization (N-pT-BN), then convolution 2 with another N-pT-BN, followed by convolution 3, and finally flattening to a 1D vector. Similarly, the second block is a 2D CNN block, where the input is processed through convolution 1, followed by N-pT-BN, then convolution 2 with N-pT-BN, and convolution 3, which is then flattened. After these operations, the outputs from both blocks are merged in the fusion layer, where the features from the 1D and 2D blocks are combined. The fused features are then passed through a series of FC layers, followed by a dropout layer, then FC, and finally the Softmax layer, which outputs the class probabilities for classification tasks. This architecture efficiently combines 1D and 2D convolutional features for comprehensive feature extraction and classification. The suggested NPCNN model involves N-pT-BN in each branch to normalize activations prior to feature fusion. Such a method helps reduce redundancy, increase feature diversity, and improve classification performance in difficult traffic situations.

The N-pT-BN process begins with the tanh activation function (Nwankpa et al., 2018), as provided in Equation 16. Then, the activation function is normalized, as provided in Equation 17, and Lp pooling p g y (Gholamalinezhad and Khosravi, 2020) is applied to the normalized activation function (Equation 18). Finally, BN (Ioffe and Szegedy, 2015) is applied to the pooled normalized activation function, as provided in Equation 19: where η g y = g y − μ σ is the normalization function and μ and σ denote the mean and standard deviation of g y, respectively. where b i is the feature rate at location i within the pooling region γ j, ε denotes constant, y i refers to the prior layer feature map, and y ^ i denotes N-pT-BN and the value of v in the range 1 < v < ∞.

g y = tanh y i , y ≥ 0 b * ⁡ tanh y i , y < 0 , (16)

g ′ y = g y * η g y , (17)

p g y = g ′ y * 1 γ j ∑ i ∈ γ j b i v 1 v , (18)

y ^ i = y i − μ σ 2 + ε * p g y , (19)

Thus, the proposed model adopts BN, which ensures stable outcomes and improves the training dynamics. It also utilizes the tanh activation function followed by a normalized derivative to enhance feature contrast and Lp pooling focus on strong activations balancing among pooling layers. The output of the object detection obtained from the NPCNN model is type labels such as “0, cyclist”; “1, person”; and “2, people.”

4 Results and discussion

4.1 Simulation procedure

The proposed object detection in traffic scenarios was simulated using PYTHON 3.7. The processor employed was Intel(R) Core (TM) i5-1035G1 CPU @ 1,00 GHz 1.19 GHz and the Installed RAM size was 20.0 GB. The analysis of object detection in traffic scenarios was conducted using datasets from KAIST (https://www.kaggle.com/datasets/adlteam/kaist-dataset?select=set00) and Cityscapes (https://www.kaggle.com/datasets/shuvoalok/cityscapes—used for cross validation only).

4.2 Dataset 1 description: KAIST dataset

The dataset was imaging hardware comprising a color camera, a thermal camera, and a beam splitter to seize the affiliated multispectral images. It captured numerous traffic locations during both day and night to reflect variations in light settings and encompassed 95k color-thermal pairs (640×480, 20 Hz) occupied from a vehicle. The pairs were annotated manually (Person, People, and Cyclist). The dense annotations included 103,128, and unique pedestrians consisted of 1,182.

For our study, 2,574 data samples were employed, divided across three different classes: Cyclist (Class 0), Person (Class 1), and People (Class 2). Each class encompassed 858 data samples.

4.2.1 Dataset 2 description: Cityscapes dataset (used for cross validation only)

The Cityscapes dataset is an extensively used benchmarking datasets for urban scene analysis and object recognition for autonomous driving applications. The dataset consists of high-quality images captured using the camera mounted on vehicles in 50 different cities across Germany and neighboring European countries. Dataset 2 is partitioned with increasing training proportions while keeping validation constant at 20%. At 40% training, the dataset consists of 2,703 training samples, 676 validation samples, and 2,253 test samples, each with 74 features. With a 50% training split, the training set expands to 3,153 samples, validation to 789 samples, and testing reduces to 1,690 samples. For 60% training, there are 3,604 training samples, 901 validation samples, and 1,127 test samples. At the highest training proportion of 70%, the dataset contains 4,054 training samples, 1,014 validation samples, and 564 test samples, maintaining the same 74 features per sample throughout. The training, validation, and testing details of the dataset are tabulated in Table 2.

TABLE 2. Training, validation, and testing details of both datasets.

4.3 Performance analysis

A comparative evaluation was conducted to investigate the efficiency of NPCNN and existing approaches for object detection in traffic scenarios. The estimation was performed across a wide range of metrics, such as accuracy, precision, sensitivity, specificity, FDR, FNR, FPR, F-Measure, MCC, and NPV. The ablation study, statistical analysis, and ROC curve evaluation were accomplished to confirm the efficacy of the NPCNN approach. In addition, the NPCNN strategy was compared with state-of-the-art models such as FSRCNN (Garcia-Aguilar et al., 2023), R-CNN (Chughtai et al., 2024), YOLOv8 (Naufaldihanif et al., 2025), and RT-DERT (Cui et al., 2025) and existing schemes such as Bi-LSTM, LinkNet, EfficientNet, SqueezeNet, and PCNN. The hyperparameters of the classifiers is shown in Table 3.

TABLE 3. Hyperparameters of the classifiers.

The NPCNN and conventional approaches were analyzed using the KAIST dataset. Figure 5 displays the sample images and their corresponding pre-processed results after applying Wiener filtering.

4.4 Segmentation analysis

Figure 6 presents a thorough overview of original images with their respective segmented outcomes through numerous segmentation techniques, including Conventional Mask RCNN, ResNet, U-Net, and MBA-MRCNN. Notably, MBA-MRCNN outperforms existing methods, affording more precise and detailed segmentation outcomes. This indicates that MBA-MRCNN demonstrates its enhanced functionality over conventional methodologies.

4.4.1 Object detection results

Figure 7 presents the qualitative object detection results of the proposed NPC-ODT framework. The left column shows the original traffic images, while the right column illustrates the corresponding detection results with bounding boxes and class labels. The proposed framework successfully detects and localizes different traffic participants, including cyclists and pedestrians, by generating accurate bounding boxes with correct class predictions. The results demonstrate that NPC-ODT effectively identifies objects under varying traffic conditions, including small object sizes and cluttered backgrounds, highlighting its robustness and suitability for real-world traffic scene analysis.

4.4.2 Assessment on Dice and Jaccard

The MBA-MRCNN model performance was analyzed using the Dice and Jaccard metrics, and comparisons were made with traditional strategies such as Conventional Mask RCNN, U-Net, and ResNet (Table 4). Examining the Dice metric, MBA-MRCNN surpassed traditional methods, accomplishing a remarkable Dice value of 0.903. In contrast, Conventional Mask RCNN, U-Net, and ResNet attained lower Dice ratings of 0.730, 0.522, and 0.652, respectively. Moreover, MBA-MRCNN reached the peak Jaccard value of 0.885, outperforming Conventional Mask RCNN, U-Net, and ResNet. The MBM helps mask RCNN focus better on important parts of images, thus improving object detection and mask accuracy.

TABLE 4. Dice and Jaccard analysis of MBA-MRCNN versus conventional methods.

4.5 Comparative analysis

A comprehensive comparative assessment was conducted against conventional methods, including FSRCNN (Garcia-Aguilar et al., 2023), R-CNN (Chughtai et al., 2024), YOLOv8 (Naufaldihanif et al., 2025), RT-DERT (Cui et al., 2025), Bi-LSTM, LinkNet, EfficienctNet, SqueezeNet, and PCNN, with the results presented in Figures 8–10. For a method to be reflected as effective in object detection, it is essential that it exhibits higher values in positive and neutral metrics values, thus demonstrating its ability to exactly recognize and localize objects within traffic environments. At 60% training data, the NPCNN obtained an impressive accuracy of 0.898, suggestively excelling the performance of existing methods. As the training data heightened to 70% and 80%, the NPCNN model continued to surpass the established models, achieving accuracy rates of 0.916 and 0.937. When the training data reached 90%, the NPCNN strategy accomplished the highest accuracy of 0.956, while conventional methods recorded relatively lower accuracy ratings with FSRCNN at 0.824 (Garcia-Aguilar et al., 2023), R-CNN at 0.887 (Chughtai et al., 2024), Bi-LSTM at 0.864, LinkNet at 0.841, EfficientNet at 0.875, SqueezeNet at 0.869, and PCNN at 0.854, respectively. For 90% training data, the NPCNN scheme exhibited a higher specificity rate of 0.964, signifying its enhanced object detection capability. In contrast, FSRCNN (Garcia-Aguilar et al., 2023), R-CNN (Chughtai et al., 2024), Bi-LSTM, LinkNet, EfficientNet, SqueezeNet, YOLOv8 (Naufaldihanif et al., 2025), RT-DERT (Cui et al., 2025), and PCNN generated lesser specificity values of 0.910 to 0.932.

When analyzing the performance of 70% training data, NPCNN acquired the maximum NPV of 0.931, surpassing traditional approaches. Among the models, FSRCNN (Garcia-Aguilar et al., 2023) recorded the lowest NPV of 0.835, while LinkNet, PCNN, and SqueezeNet generated NPV values of 0.865, 0.874, and 0.871. R-CNN (Chughtai et al., 2024) and Bi-LSTM achieved NPV of 0.915. Lesser negative metric scores are essential for efficient object detection, signifying the efficacy of the model in lessening errors and advancing detection accuracy. Considering the FDR metric, the NPCNN scheme achieved the lowest FDR rate of 0.114 at 60% training data, while FSRCNN (Garcia-Aguilar et al., 2023), R-CNN (Chughtai et al., 2024), Bi-LSTM, LinkNet, EfficientNet, SqueezeNet, PCNN, YOLOv8 (Naufaldihanif et al., 2025), and RT-DERT (Cui et al., 2025), demonstrated higher FDR values of 0.287, 0.201, 0.241, 0.43, 0.185, 0.239, and 0.198, respectively. Finally, at 90% training data, the NPCNN approach generated minimized FDR values of 0.057. Comparatively, traditional strategies such as FSRCNN (Garcia-Aguilar et al., 2023), R-CNN (Chughtai et al., 2024), Bi-LSTM, LinkNet, EfficientNet, SqueezeNet, and PCNN scored higher FDR ratings. UC-HoS updates the center pixel before skeletonization helps reduce noise and enhance edge definition, resulting in a cleaner image. This leads to a more accurate and robust skeletonization process and a better foundation.

4.6 Ablation study

By systematically removing components, the ablation study emphasized the individual impact of each module within the overall NPC-ODT approach. Table 5 describes the ablation evaluation of NPC-ODT for object detection in traffic conditions. This analysis contrasts the NPC-ODT model against several variants, such as NPC-ODT with existing feature extraction, NPC-ODT with existing segmentation, NPC-ODT without segmentation, NPC-ODT without feature extraction, and NPC-ODT with existing classifier. The NPC-ODT model attained the lowest FPR rate of 0.047, exhibiting its robustness in object detection for traffic scenarios. Comparatively, NPC-ODT with existing classifier and NPC-ODT with existing segmentation scored respective maximum FPR scores of 0.077 and 0.079. NPC-ODT without segmentation and NPC-ODT excluding feature extraction yielded respective FPRs of 0.107 and 0.106. NPC-ODT with existing feature extraction achieved the maximum FPR rate of 0.128. In terms of the F-measure, the NPC-ODT scheme accomplished the maximum score of 0.906. In comparison, the NPC-ODT employing existing feature extraction, NPC-ODT utilizing existing segmentation, NPC-ODT without segmentation, NPC-ODT excluding feature extraction, and NPC-ODT employing existing classifier generated lower F-measure values.

TABLE 5. Ablation analysis on NPC-ODT strategy, NPC-ODT with existing feature extraction, NPC-ODT utilizing existing segmentation, NPC-ODT without segmentation, NPC-ODT excluding feature extraction, and NPC-ODT employing existing classifier.

4.6.1 Ablation study analysis of individual component

In Table 6, the proposed NPC-ODT framework’s performance analysis has been performed. It evaluates the effectiveness of individual proposed components such as MBA-based segmentation features, UC-HoS feature extraction, and N-pT-BN normalization technique, along with the complete proposed integrated framework. The proposed model with MBA-based segmentation features yields the lowest performance for the proposed variant models, with an accuracy of 0.842, sensitivity of 0.763, specificity of 0.882, precision of 0.765, F-measure of 0.764, and MCC of 0.646. Additionally, it achieves the highest error rates, with an FPR of 0.117, FNR of 0.236, and FDR of 0.234. The use of the proposed UC-HoS feature extraction yields better performance results, with accuracy of 0.876, sensitivity of 0.813, precision of 0.815, F-measure of 0.814, and MCC of 0.721. It also obtains the minimum error values, with FPR of 0.046, FNR of 0.094, and FDR of 0.093. This shows that the proposed normalization technique is effective enough to stabilize feature learning, network convergence, and classifier discriminative ability.

TABLE 6. Analysis of proposed NPC-ODT framework using individual proposed components.

4.7 Statistical analysis on accuracy

Table 7 presents an exhaustive statistical assessment of the NPCNN approach, contrasting its performance against existing strategies, including FSRCNN (Garcia-Aguilar et al., 2023), R-CNN (Chughtai et al., 2024), Bi-LSTM, LinkNet, EfficientNet, SqueezeNet, and PCNN for object detection in traffic scenarios. The maximum statistical measure, NPCNN, had the highest accuracy value of 0.957, while the conventional approaches had lower accuracy values of 0.872–0.932. In addition, the highest accuracy rate acquired using the NPCNN model is 0.928 in mean statistical metric, while the FSRCNN (Garcia-Aguilar et al., 2023), R-CNN (Chughtai et al., 2024), Bi-LSTM, LinkNet, EfficientNet, SqueezeNet, and PCNN generated the lowest accuracies of 0.844, 0.896, 0.875, 0.863, 0.907, 0.877, and 0.890, respectively. This emphasizes the enhanced performance and reliability of NPCNN in traffic scenario object detection.

TABLE 7. Statistical evaluation of accuracy.

4.8 Mean average precision comparison of NPCNN and existing models

The comparison of the mean average precision (mAP) obtained from the proposed NPCNN approach and several existing object detection algorithms for four different training data ratios (60%, 70%, 80%, and 90%) is illustrated in Figure 11. The proposed NPCNN model always performs better than other models, regardless of the ratio of training data used. This is because the mAP obtained from the NPCNN model is 0.8800, 0.8800, 0.8802, and 0.8807 for 60%, 70%, 80%, and 90% training data, respectively—much higher than the mAP obtained from the compared algorithms. In terms of the existing algorithms, SegNeXt and EfficientDet have good performance with maximum mAP of 0.8470 and 0.8462, respectively, while others such as YOLOv8, RT-DETR, FSRCNN, R-CNN, Bi-LSTM, and PCNN have mAP in the range of 0.8430–0.8474. This demonstrates that the proposed model is quite robust and has a good generalization capability.

4.9 Cross dataset validation analysis

The cross-dataset analysis indicates that the suggested model consistently performs well, implying strong generalizability of the model across various traffic datasets. The cross-validation results of both datasets are tabulated in Table 8. Specifically, when trained using Dataset 1 and evaluated using Dataset 2, the model produced an accuracy of 93.5%. Similarly, when the model was trained using Dataset 2 and evaluated using Dataset 1, its accuracy improved slightly to 93.6%. Likewise, the sensitivity of the model was enhanced from 92.7% to 93.6%, while the specificity stayed almost constant at 92.3% and 92.5%. Therefore, the suggested approach shows good generalizability across different datasets, providing high accuracy, precision, and stable classification results.

TABLE 8. Cross validation analysis of both datasets.

4.10 Analysis on ROC curve

A graphical depiction marks the TPR against the FPR at several threshold. This analysis on the NPCNN approach in comparison to conventional methods such as FSRCNN (Garcia-Aguilar et al., 2023), R-CNN (Chughtai et al., 2024), Bi-LSTM, LinkNet, EfficientNet, SqueezeNet and YOLOv8 (Naufaldihanif et al., 2025), RT-DERT (Cui et al., 2025), and PCNN for object detection in traffic scenarios is displayed in Figure 12. The outcomes of the ROC curve analysis show that the NPCNN approach surpasses all conventional methods in terms of AUC score. The NPCNN scheme achieved an impressive AUC rate of 0.92, significantly outperforming R-CNN (Chughtai et al., 2024) with an AUC rate of 0.90, showcasing its outstanding performance. While R-CNN (Chughtai et al., 2024) also presented robust performance, the NPCNN approach was shown to be more efficacious. Bi-LSTM and LinkNet attained an AUC score of 0.86, though FSRCNN (Garcia-Aguilar et al., 2023), EfficientNet, SqueezeNet, and PCNN recorded AUC values of 0.83, 0.88, 0.87, and 0.89. These models exhibited excellent performance but did not outperform NPCNN. Lp pooling in PCNN softly focuses on strong activations balancing between pooling layers. Batch normalization ensures stable outputs and enhances the training dynamic, making the layer both selective and robust.

4.11 Analysis of computational complexity using FLOPS

Table 9 compares the computational complexity of different classifiers using FLOPS (GFLOPS). Among the evaluated methods, R-CNN exhibits the lowest computational complexity with 9.1052 GFLOPS, followed by BiLSTM (9.3743 GFLOPS) and FSRCNN (9.7174 GFLOPS). The proposed model requires 10.217 GFLOPS, which is lower than PCNN (11.084 GFLOPS), RT-DETR (11.334 GFLOPS), EfficientNet (11.711 GFLOPS), LinkNet (12.452 GFLOPS), YOLOv8 (12.806 GFLOPS), and SqueezeNet (13.159 GFLOPS). These results indicate that the proposed model achieves a favorable balance between computational efficiency and detection capability.

TABLE 9. Analysis of FLOPS.

5 Conclusion

This study proposes a novel normalized parallel convolutional neural network for object detection in traffic scenarios (NPC-ODT). The process of object detection in traffic scenarios began with an input video, which was first divided into individual frames for detailed analysis. Each frame underwent preprocessing using the Wiener filter, which helped reduce noise while preserving important image details, thereby enhancing the quality of input data. Following this, segmentation was performed using a MBA-MRCNN, which effectively isolated regions of interest such as pedestrians, vehicles, and road signs from the background. Once the relevant segments were identified, feature extraction was performed based on a hierarchy of methods. This included UC-HoS that captured shape-related cues, color-based features for object differentiation, and motion estimation to track moving elements across frames. These extracted features were then fed into a NPCNN, which was trained to detect and classify objects within the scene. The final output consisted of accurately detected and labeled objects in each frame, enabling further tasks such as tracking or behavior analysis. When analyzing the performance of 70% of the training data, NPCNN attained the maximum NPV rate of 0.931, outperforming traditional approaches. Among the models, FSRCNN (Garcia-Aguilar et al., 2023) recorded the lowest NPV of 0.835, while LinkNet, PCNN and SqueezeNet generated NPV values of 0.865, 0.874, and 0.871, respectively.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Glossary

AFFB Attention feature fusion block AUG Augmentation technique BM Block matching BN Batch normalization Bi-LSTM Bidirectional long short-term memory CBL Convolution-Batchnorm-Leaky ReLU CNN Convolutional neural network DRUformer Driving scene relationship understanding transformer DSSD Deconvolutional single-shot detector FC Fully connected FSA Full Search algorithm FSRCNN Fast super-resolution convolutional neural network model FTOD Fogged traffic image dataset GAN Generative adversarial network HRDNet High-resolution detection network LSKNet Large selective kernel network MBA Modified bottleneck attention MBA-MRCNN Modified bottleneck attention-based mask recurrent convolutional neural network MFEF Multi-scale feature enhancement fusion MS-CAB Multi-scale channel attention block NPCNN Normalized parallel convolutional neural network NPC-ODT Normalized parallel convolutional neural network for object detection in traffic scenarios N-pT-BN Normalized pTanh response batch normalization OCIR Online instance classifier refinement PAE Portable appearance extension PCNN Parallel convolutional neural network PSNR Peak signal-to-noise ratio ReLU Rectified linear unit R-CNN Recurrent convolutional neural network RepViT Revisiting perspective vision transformer RoI Regions of interest RPN Region Proposal Network SAD Sum of absolute differences SCSA Shuffle channel and spatial attention SGBlock Sand glass block SNIP Scale normalization for image pyramids SOTA State-of-the-Art SPD-Conv Space-to-depth convolution UC-HoS Updated center pixel in hierarchy of skeleton Z-YOLOv8s Z-You Only Look Once version 8 small

Footnotes

Edited by: Roberto Ventura, University of Brescia, Italy

Reviewed by: Govindhan Padmapriya, SRM Institute of Science and Technology (Deemed to be University) Research Kattankulathur, India

Hongjie He, Zhengzhou University of Light Industry, China