Optimizing drivable area detection for autonomous vehicles through shadow and road edge line analysis in U-net-based models

1 Introduction

Driver behavior accounts for an estimated 94% of vehicular accidents. This statistic underscores the fundamental motivation behind intelligent connected vehicle development: collision prevention through advanced automation (Singh, 2015). The rapid deployment of autonomous driving systems has been accompanied by heightened safety concerns, particularly regarding collision incidents involving automated vehicles. The security of autonomous driving systems depends on various factors, including the vehicle operating system (Wu, 2021; Cheng et al., 2022), environmental perception (Cui et al., 2021; Guan et al., 2023), high-precision mapping and positioning (Cheng et al., 2022; Guan et al., 2024), predictive decision-making, and planning, control (Huang and Chen, 2020), and execution modules (Maurer et al., 2016). Drivable area detection is a crucial component of environmental perception and predictive decision-making. The term “drivable area” refers to a road region that can be delineated at the pixel level, enabling vehicles to navigate safely and normally within the defined boundaries (Shinzato and Wolf, 2011). Drivable area detection is critical for autonomous driving safety, as it provides essential information, such as road boundaries and obstacles. However, Urban roads present complex driving environments, incorporating diverse elements such as lane markings, obstacles, and traffic signs. Furthermore, these environmental elements exhibit significant variability under changing weather and illumination conditions, adversely affecting detection algorithm accuracy. The inherent complexity and dynamic nature of urban traffic environments pose substantial challenges for autonomous vehicles in reliably identifying drivable areas (Liu et al., 2018).

The accuracy of autonomous driving detection depends on two primary factors. The first is the integrity of the main system of autonomous driving technology. Enhanced detection accuracy can be achieved by optimizing the methods used to detect the driving area. The second is road infrastructure improvement, which allows for effective recognition by autonomous vehicles, further enhancing detection accuracy. Extensive research has been conducted to enhance the autonomous recognition and judgment of self-driving vehicles during road travel. Such research has involved the construction of models and methodologies. Current studies primarily utilize techniques, such as common point mapping (Liu et al., 2018), YOLO-based segmentation algorithms (Wang et al., 2024), camera-based machine vision segmentation algorithms (Neto et al., 2013), and image processing technologies (Devane et al., 2021), to facilitate driving area detection. The advantages of these methods lie in their ability to significantly enhance the vehicle’s perception accuracy of the road environment. For example, the YOLO - based segmentation algorithm, with its efficient real-time object detection capability, substantially reduces the misjudgment rate in complex scenarios. Meanwhile, the co-point mapping technology effectively addresses the limitations of single sensors in dynamic environments through multi - modal data fusion. However, existing methods are predominantly designed from the perspective of ensuring primary vehicle system safety. While these methods enhance drivable area detection accuracy for autonomous vehicles, they encounter critical challenges. Current systems exhibit excessive dependence on ego-vehicle perception capabilities while demonstrating limited adaptability to built environment influences. Particularly in scenarios featuring shadow occlusion and low chromatic contrast between drivable and non-drivable surfaces, the algorithms’ robustness remains inadequate. In fact, the built environment’s effect on drivable area determination is non-negligible. Fraedrich et al. (2019) reported that fully autonomous driving is likely to affect the built environment in the future, thereby intersecting with the domain of urban planning. The Royal Automobile Club Foundation has identified serious shortcomings in existing road infrastructure with regard to its capacity to support connected and autonomous vehicles. The Foundation emphasizes the need for improvements and updates in road markings, traffic signals, pavement conditions, and communication networks to accommodate these emerging technologies.

Nevertheless, significant challenges persist in developing intelligent infrastructure with seamless compatibility for autonomous vehicle integration (Fraedrich et al., 2019). The majority of contemporary research that examined the interactions between the built environment and autonomous driving has focused on the effect of automated driving on the built environment and land use. A study conducted by Florida State University (Chapin et al., 2016) investigated the manner through which the built environment evolves in response to the advent of self-driving vehicles. The study’s key findings indicate that the introduction of self-driving cars requires narrow right-of-way and travel lanes; influence the location, design, and quantity of parking spaces; affect the mobility of cyclists and pedestrians; contribute to a tidy urban environment by reducing the need for signals and signage; and enable the redevelopment of currently unnecessary parking lots and redundant rights of way. Vaddadi (2017) argues that autonomous driving systems could enhance urban traffic coordination through smart parking solutions and similar innovations. Such systems may alleviate parking-related challenges while reducing land dedicated to parking infrastructure, potentially transforming urban spaces and enabling alternative land use. However, these studies exhibit a critical limitation: they predominantly examine how autonomous vehicles might reshape the built environment, neglecting the reciprocal constraints imposed by existing urban infrastructure on autonomous driving performance. For example, irregular road designs may lead to vehicle positioning errors (Mihalj et al., 2022). The impact of autonomous driving on the built environment is not unidirectional. As this relationship evolves, we must also examine how the built environment shapes the performance and adoption of autonomous driving technology.

The accuracy of drivable area detection is significantly constrained by two key urban environmental factors: (1) road shadows cast by vegetation and buildings (Yu and Marinov 2020), and (2) road edge markings (Park and Jayachandran Raguraman, 2020). Current research has yet to establish quantitative measures of these interference effects or develop corresponding optimization approaches for autonomous vehicle perception systems in built environments. This study implements a deep learning-based semantic segmentation model for road scenes to evaluate how shadows and road edge lines affect the accuracy of drivable area detection. A U-Net-based model for semantic segmentation of road scenes is used in this study. Compared to other deep learning architectures, U-Net’s skip connections enable superior detail recovery in segmentation tasks. The skip connections in U-Net facilitate the fusion of low-level encoder features with high-level decoder features, enabling precise boundary recovery in segmentation tasks. This capability is particularly crucial for accurately segmenting small objects and complex structures in street scenes, where fine details significantly impact segmentation quality (Ronneberger et al., 2015). U-Net has been widely employed in medical imaging, satellite imagery, and street scene segmentation, and its variants and extensions (e.g., U-Net++ and Attention U-Net) exhibit further improved performance and flexibility (Oktay et al., 2018).

This study aims to analyze how the built environment affects drivable area segmentation for autonomous driving. We propose an improved U-Net model, train and optimize it via convolutional neural networks, and validate its performance on drivable area segmentation tasks. Eventually, the research conclusion will be to evaluate the impact of road shadows and road boundary lines on the accuracy of semantic segmentation of the street scenes for autonomous driving. Two innovative research approaches proposed in this study. The first innovative research approach is to start from the perspective of the built environment, study the impact of the built environment on the driving safety of autonomous vehicles on the road, and explain the specific mechanism by which the elements of the built environment affect the identification of the drivable area by autonomous driving. The second one is to introduce a deep learning model to simulate the research on the identification of the drivable rea in autonomous driving.

2 Literature review

2.1 Drivable area for autonomous driving (autonomous vehicle safety system)

Subdividing the driving area for vehicles provides environmental awareness to autonomous vehicles and improves the system’s understanding of safe driving zones. In the Defense Advanced Research Projects Agency Competition, Dolgov et al. (2009) proposed a method that uses LiDAR sensors to extract road positions and combines radar obstacle detection functionality to obtain information about obstacles around the vehicle, thereby identifying the 3D space in which the vehicle is allowed to drive. This method uses the height differences in the point cloud data collected by LiDAR to detect the road surface; removes areas classified as sky and obstacles to accurately identify the drivable road sections, allowing for improved navigation; and constructs a 3D view of the drivable area. Fan et al. (2020) proposed a new method called Sne-roads, which improves the segmentation effect of free space (i.e., drivable area) by fusing normal information on the surface. This method can accurately identify and segment the drivable road area in complex road scenes, especially when it is faced with challenging environments, such as different ground features and surface changes. Rasmussen (2002) combined depth information from laser rangefinders with color texture information from visual sensors to train a neural network model. A large number of registered laser and camera images were captured at frame rate on different rural roads. In addition, 3D laser height and smoothness features were associated with color histograms and bandpass filter response image features, a set of road models suitable for path planning was generated by training separate neural networks on tagged feature vectors clustered by road type, and a 3D map suitable for path planning was constructed. Since the proposal of deep learning, CNN-based deep learning algorithms have been used to identify drivable areas for autonomous driving. Deep learning models for semantic segmentation include U-Net (Ronneberger et al., 2015), fully convolutional network (Long et al., 2015), SegNet (Badrinarayanan et al., 2017), segment anything model (SAM) (Dave et al., 2019), and fast SAM (Zhao et al., 2023). Rawashdeh et al. (2023) investigated the use of CNNs for the integration of forward-looking cameras with LiDAR sensors for the detection of driving areas or roads. Several CNN architectures for sensor fusion were compared in terms of accuracy, runtime, and size. To address sensor fusion challenges in autonomous driving, Li et al. (2022) introduced a Transformer-based fusion framework (TransFusion) that combines camera and LiDAR data for robust 3D object detection. This approach establishes new state-of-the-art performance in adverse weather conditions by leveraging cross-modal attention mechanisms to dynamically align features and mitigate calibration errors 15. In parallel, Lang and Wang (Lang and Wang, 2024) proposed ICAFusion, an iterative cross-attention guided framework for multispectral object detection. Their method employs dual-branch CNNs (e.g., ResNet or CSPDarkNet) to extract RGB and thermal infrared features, followed by a lightweight transformer decoder with parameter-sharing iterative blocks. This design adaptively fuses multimodal features through spatial-aware attention, achieving a 45 FPS inference speed while improving mAP@0.5 by 4.2% on the KAIST dataset and reducing false detections by 31% in nighttime scenarios on FLIR.

2.2 Effect of the built environment (road shadows and road edge lines) on the drivable area

The interaction between autonomous driving and the built environment is a complex, dynamic process. As autonomous vehicles become increasingly popular, urban planners may need to redesign roads to make them safe for autonomous driving. They also need to consider what type of built environment is suitable for autonomous driving. Duarte and Ratti (2018) reported that if autonomous vehicles are viewed as autonomous platforms with multiple combined functions, new road and urban designs may be needed. Autonomous driving is expected to gradually change the built environment during the evolution process, which means that the current built environment is unsuitable for autonomous driving in many aspects. Therefore, from the perspective of the built environment, the effects of some current built environment elements on autonomous driving must be investigated to expand the application of autonomous driving. Since the introduction of autonomous driving, researchers have been investigating how the built environment, including land use (Bridgelall and Stubbing, 2021), road infrastructure (Othman, 2021), parking, and road markings, affects the safety of autonomous driving.

Road shadows and road edge lines influence the quality of road streetscape images in daily life. Wu et al. (2014) highlighted the importance of extracting lane markings in complex environments (e.g., shadows and ambiguous road edge lines). The two key parameters of the built environment are selected in this study.

2.2.1 Road shadows caused by roadside greenery and buildings

Liu et al. (2021b), Fang et al. (2019), and Tatar et al. (2018) found that road shadows can affect road recognition and detection. Wang et al. (2020) reported that road shadows can blur and alter road features, considerably degrade image quality, and make further detection and tracking difficult. Arnold et al. (Arnold et al., 2019) suggested that shadows can obscure object features, reduce the reliability of sensors in complex environments, and affect the performance of visual sensors. These factors can lead to detection errors or missed detections, ultimately affecting the perception and avoidance of road obstacles by autonomous vehicles. In addition, shadows cast by buildings and trees can obscure road signs and markings, making it difficult for autonomous driving systems to accurately identify important information, such as speed limits, turning points, and lane boundaries at high speeds. This situation can affect the vehicle’s heading control and safe driving decisions. Road shadows also influence the autonomous system’s detection and judgment of drivable areas (Pavel et al., 2022).

2.2.2 Road edge lines

Road edge lines have a direct substantial effect on the detection of the drivable area by autonomous driving systems, and the absence of the road edge lines makes drivable area detection difficult (Zhao et al., 2023). Kong et al. (2010) discovered that road edge lines can affect road detection, such as the drivable area, and detection can go wrong because of incorrect recognition of shadow boundaries. Adaptability and accuracy are also affected. Yan et al. (2016) reported that with the commercialization of advanced driving assistance systems and autonomous driving technology, the demand for detailed 3D road maps is expected to increase, and road markings and road edge lines are essential information for generating these maps. They proposed an innovative road marker extraction method based on scanning lines; it accurately extracts road marker information by processing LiDAR point cloud data to improve the road marker detection capability of autonomous driving systems.

3 Methods

In this section, the basic principles of U-Net for semantic image segmentation are briefly described. The modifications to the model used in this research are presented based on the U-Net architecture. The main modifications include an increased dropout rate, an increase in the number of convolutional kernels, and adjustments to the pooling layers. The main aim is to improve the robustness and generalizability of the model while reducing overfitting.

3.1 U-net

U-Net is a CNN specifically designed for image segmentation. Its basic architecture consists of two symmetric paths that form an encoder contraction path and a decoder expansion path, which are connected to the encoder and decoder by skip connections. Inter-transmitting feature diagrams allow the network to effectively use low-level features and combine high-level semantic information to improve the accuracy and details of segmentation. Specifically, in the decoder phase, the characteristic diagram of the encoder phase is stitched with the characteristic diagram of the decoder stage. This design concept originates from the fully convolutional network (Long et al., 2015) and has been improved to effectively handle fine-grained segmentation tasks.

The encoder (Sutskever et al., 2014) section consists mainly of a series of convolution and pooling layers. The convolution layer is responsible for extracting the features of the image layer by layer. Each convolution operation usually uses small convolution kernels, such as 3 × 3, and feature mapping through nonlinear activation functions, such as RELU. The pooling layer typically adopts max pooling the largest pooling Max Pooling to reduce the spatial resolution of the feature map while increasing its depth to capture advanced semantic information in the image. This process is gradually sampled down to downsampling to allow the network to extract abundant abstract features.

Decoders (Sutskever et al., 2014) restore the feature maps extracted from the encoder to the original image spatial resolution through upsampling operations. The upsampling process typically uses transpose convolution or interpolation techniques to expand the spatial dimensions of the feature maps and recover image details. After each upsampling operation, the decoder applies convolutional layers to further refine the feature maps and improve the accuracy of image segmentation. U-Net was originally designed for medical image segmentation tasks, but its effectiveness has led to its widespread application in many fields (Ronneberger et al., 2015). In satellite image analysis, it is employed to extract land cover categories, such as buildings, roads, and vegetation, from satellite imagery. U-Net also demonstrates significant value in remote sensing image analysis and geographic information systems (Pan et al., 2020). In street view image segmentation, its application spans autonomous driving and urban planning, enabling precise delineation of key urban elements—including roads, vehicles, and pedestrians. This capability is essential for high-precision environmental mapping and real-time scene understanding, forming the foundation for robust spatial analysis and decision-making in smart city applications (Kim et al., 2020).

3.2 Model refinement

This study employs an enhanced U-Net variant that preserves the fundamental encoder-decoder structure of the original architecture while incorporating targeted modifications.

3.2.1 Upsampling procedure in the expansive path

In the conventional U-Net architecture, the upsampling procedure in the expansive path is typically implemented using the Conv2DTranspose layer (also referred to as the deconvolution layer). The model used in this research introduces modifications in this regard by incorporating a combination of Conv2DTranspose layers and concatenation operations while employing different upsampling strategies and convolution kernel sizes. For example, the U6 layer uses a (3,2) convolution kernel with a (3,2) step for deconvolution, with slight variations in upsampling methodology and convolution kernel configuration compared with the original U-Net. Compared with the standardized upsampling steps in traditional U-Net, the upsampling operations in this research model are tailored at different stages to accommodate different feature map dimensions.

3.2.2 Modification of the pooling layer

The model employed in this study utilizes the pooling layer. In the standard U-Net architecture, the pooling operation typically utilizes MaxPooling2D with a pool size of (2,2). By contrast, the model presented in this study incorporates some pooling operations that use MaxPooling2D with a pool size of (3,2). This modification aims to enhance control over the spatial dimensions of the feature map, thereby enabling the network to capture detailed contextual information during the sampling process while mitigating excessive reductions in the spatial resolution of the feature map. The (3,2) pooling kernel is designed to better preserve the aspect ratio of road features, which typically extend longer in the horizontal direction, thereby reducing information loss during downsampling.

3.2.3 Configuration of convolutional kernels and strides

The model uses nonstandard configurations of convolutional kernels and strides in several convolutional layers. For example, the deconvolution layers used in U6 and U7 have nonstandard strides of (3,2) and (2,2), respectively. The purpose of this adjustment is to match the spatial resolution of the feature maps, ensuring that their dimensions closely match those of the feature maps from the corresponding downsampling stages. This configuration can improve the accuracy of information fusion within the network.

3.2.4 Utilization of the dropout layer

In the model, the dropout layer is incorporated into the expansion path, especially after convolutional operations, to mitigate the risk of overfitting,as shown in Figure 1. Compared with the traditional U-Net architecture that does not consistently incorporate a dropout layer after each convolutional operation, the model presented in this study employs different dropout probabilities, such as 0.2, 0.3, and 0.4, across different convolutional blocks. This approach enhances the regularization of the network, thereby improving its generalization capabilities during the training process. The increasing dropout rates (0.2, 0.3, 0.4) along the decoder path are set to progressively regularize deeper layers, which contain more abstract features and are more prone to overfitting given the limited training data.

3.2.5 Feature fusion and splicing

The model integrates upsampled feature maps with their corresponding downsampled feature maps through concatenation operations in the expansion path. The model precisely aligns the channel numbers of the feature maps during these operations to ensure effective feature fusion. Although this approach is consistent with the classical U-Net architecture, the specific implementation has subtle differences, such as variations in feature map sizes and channel configurations across layers.

Reducing the initial number of channels from 64 to 32 lowers model capacity at the early stage, forcing the network to learn more efficient representations while reducing the total parameter count, which mitigates overfitting and accelerates training.

In summary, the improvements to the U-Net architecture used in this research include adjustments to the operations in the pooling layer, the stride and kernel size in the upsampling layer, different dropout strategies, and feature map alignment and concatenation at the feature fusion stage. The main goals of these improvements are to increase the expressiveness of the model, improve the accuracy of feature map alignment, and effectively control the training process of the model. The overall architecture of the modified model is shown in Figure 2. Model architecture improvements and superiorities over traditional U-Net are presented in Table 1.

TABLE 1. Improvements in model architecture and superiorities over traditional U-net.

4 Training implementation model training

4.1 Model training

The training of the improved U-Net-based model is straightforward. First, the image data and the corresponding mask data are loaded using the Camvid dataset (Brostow et al., 2009), preprocessed, and saved. Second, the number of categories present in the mask data is calculated and outputted, and the dimensions and number of channels of the image are normalized. Third, image features are extracted and compressed in stages through convolutional and pooling layers (contraction path). Each stage of the convolutional layer includes parameters, such as activation function, weight initialization, and padding. Last, pooling layers are used to reduce the dimensionality of the feature maps, the model architecture is shown in Figure 3.

The load_data function is responsible for loading the image and its corresponding mask from the specified directory, followed by resizing to a uniform dimension. Subsequently, the image and mask data are stored in the X_train and Y_train arrays, respectively, to facilitate model training. The input layer of the U-Net model is designed to accept input images with dimensions (IMG_HEIGHT: 480, IMG_WIDTH: 360, IMG_CHANNELS: 3). A lambda layer is utilized to normalize the pixel values to the range [0, 1] for guaranteeing consistent scaling and input data stability. The contracting path comprises multiple convolutional layers (Conv2D) and pooling layers (MaxPooling2D), which progressively extract features from the input image. Each convolutional layer is immediately succeeded by a max pooling layer, which reduces the spatial dimensions of the feature maps while maintaining the most salient features. Regularization is applied after each downsampling step via a dropout layer to mitigate the issue of overfitting. Subsequently, the expansive path employs an inverse convolution layer (Conv2DTranspose) to upsample the feature maps and gradually restore the spatial resolution of the image. After each upsampling operation, feature maps from the contracting path are concatenated to leverage high-resolution information. Additionally, convolutional layers (Conv2D) and dropout layers are integrated within this path to enhance the model’s generalizability.

In the expansion path, transpose convolutional layers are employed to restore the compressed feature maps to their original image size and connect them to the corresponding layers in the contraction path. A series of convolutional and dropout layers are reapplied to optimize and extract features further. Then, the processed image results are output. The core evaluation metrics used to quantify segmentation performance are defined in Equations 1–5, covering loss function, class-wise IoU, mean IoU, and F1 score for pixel-level semantic segmentation assessment.

4.1.1 Calculation of loss values and accuracy rates

During the training process, U-Net employs categorical cross-entropy loss (Krizhevsky et al., 2017) to optimize the model parameters. By minimizing the loss function, the network gradually learns to classify different regions of the image accurately in the segmentation task, thus achieving accurate segmentation results.

During the training process, the model assesses the efficacy of the current parameter settings by calculating the loss value. The loss value is determined by comparing the model’s predictions with the actual labels through the formula is (Krizhevsky et al., 2017):where C represents the number of categories, yi denotes the true label, and pi signifies the predicted probability. Furthermore, accuracy is calculated based on the consistency between the model’s predicted results and the actual labels, and it reflects the performance of the model in a specific task. The formula for accuracy is (Long et al., 2015):

l o s s = − ∑ i = 1 C y i ⁡ log p i (1)

Accuracy = NumberofCorrectPixels TotalNumberofPixels (2)

This formula represents the proportion of correctly predicted pixels as a percentage of the total number of pixels.

While accuracy is intuitive, it can be misleading when class distributions are imbalanced. Therefore, we also employ the mean Intersection over Union (mIoU), which is widely recognized as the primary evaluation metric for semantic segmentation. For each class c, the IoU is computed as:where TP _*c*,FP _*c*, and FN _*c* denote true positives, false positives, and false negatives for class c,respectively. The mIoU is then the average of IoU over all C classes:

I o U c = T P c T P c + F P c + F N c (3)

m I o U = 1 C ∑ c = 1 C I o U c (4)

The F1 score (also known as Dice coefficient) is the harmonic mean of precision and recall. It can be directly derived from the IoU by the relationship:

F 1 c = 2 × I o U c 1 + I o U c (5)

The overall F1 score reported in this paper is the macro-average over all classes, i.e., the average of FAll reported results are obtained on the test split of the CamVid dataset using the best-performing checkpoint of each model during training. The same evaluation protocol is applied to all compared models to ensure fairness.

4.2 Model accuracy and loss

The model is trained using the Camvid dataset. The training process is iterated for each image file and mask file to obtain the loss value and accuracy of the model, as illustrated in Figure 4. The example images are shown in Figure 5. The comparison between the accuracy of the modified UNet model and that of traditional models is shown in Table 2.

TABLE 2. Comparison of the accuracy of the modified UNet network and that of traditional semantic segmentation models on the CamVid dataset.

However, the comparison presented in Table 2 involves different datasets (Pascal VOC 2012; CamVid, ADE20k), which may introduce dataset bias and hinder a fair evaluation of model performance. Anilkumar and Venugopal (2022) state that the SegNet model achieves a mean Intersection over Union (mIoU) of 60.1% on the CamVid dataset, while the FCN model (FCN8) achieves an mIoU of 57.0% on the same dataset. We summarize the quantitative results of the Intersection over Union (IoU) metric for each model on the CamVid dataset in Table 3. This unified comparison demonstrates that our modified U-Net consistently outperforms the baselines on the same benchmark, thereby validating the effectiveness of our improvements.

TABLE 3. Fair comparison of different models on the CamVid dataset.

4.3 Ablation experiments

To validate the contribution of the three core improvements proposed in this paper, we conducted subtraction-based ablation experiments. Starting from the full model, we sequentially removed the following key improvements:

• A (Dynamic pooling): Reverting the pooling size of certain MaxPooling2D layers from (3,2) to (2,2) and the strides of the corresponding upsampling Conv2DTranspose layers from (3,2) to (2,2);
• B (Hierarchical Dropout): Removing all Dropout layers in the decoder path (including rates 0.2, 0.3, and 0.4);
• C (Reduced initial channels): Restoring the number of filters in the first convolutional layer from 32 to 64 (i.e., changing kernel_cores from [32,64,128,256,512] back to the standard U-Net [64,128,256,512,1024]).

All four model variants (Full Model, Full-A, Full-B, Full-C) were trained and evaluated on the same CamVid dataset under identical settings (input size 480 × 360, optimizer Adam, loss sparse_categorical_crossentropy, early stopping, and model checkpoint). The quantitative results are summarized in Table 3.

As shown in Table 3, removing any single improvement leads to a drop in both accuracy and mIoU. Removing C (reduced initial channels) causes the most substantial increase in parameter count (from approximately 18.5M–31.0M) along with a clear decrease in accuracy. Removing B (Dropout) results in overfitting (high training accuracy but lower test accuracy). Removing A (dynamic pooling) degrades the model’s ability to segment fine boundaries. These results demonstrate that each of the proposed improvements contributes effectively to the drivable area detection performance.

5 Experiments

5.1 Datasets

We evaluate our models on two datasets. The CamVid dataset (Brostow et al., 2009) consists of 701 images (367 for training, 100 for validation, and 234 for testing) with a resolution of 480 × 360 pixels, annotated with 11 semantic classes. We use the standard training/validation/test split provided by the dataset. Our self-collected dataset contains 119 RGB images captured in Chongqing, China, under daytime conditions with varying shadow intensities, manually annotated using LabelMe. After data augmentation (random rotation, horizontal flipping, and brightness adjustment), the 119 original images were expanded to 476 images, which served as model input for the built-environment analysis.

5.2 Implementation details

All models are implemented in TensorFlow 2. x (Keras API) and trained on an NVIDIA Tesla T4 GPU (16 GB memory). Input images are resized to 480 × 360 and normalized to the range [0,1] via a Lambda layer. The optimizer is Adam with an initial learning rate of 1 × 10⁻⁴. A batch size of 1 is used due to memory constraints. Early stopping with a patience of 10 epochs (monitoring validation loss) is applied, and the model checkpoint saving the best validation accuracy is retained. The maximum number of epochs is set to 500, but early stopping typically reduces training to approximately 150–200 epochs.

5.3 Built environment variable selection

The effect of the built environment on the capability of an automated driving system to identify locations suitable for navigation is substantial. A number of environmental factors impede the enhancement of recognition accuracy. Among the most prevalent influences are shadows (Yu and Marinov 2020) and road edge lines (Park and Jayachandran Raguraman, 2020). In particular, shadows can lead to issues, such as reduced image contrast and blurred boundaries. Yu et al. emphasized the necessity of considering light variations and shadows when developing lane detection and obstacle recognition systems. Similarly, Raguraman et al. highlighted the importance of road edge lines in identifying drivable areas. Edge lines are critical features that help the system delineate the boundaries of the road, thereby enhancing detection accuracy and safety. In this study, a total of 968 images from the CamVid dataset were used during the model training process. Our self-collected data were collected in Chongqing City, and a total of 119 images of shadow contrast maps and road edge lines were selected. After data processing, the model input contained a total of 476 images.

5.3.1 Shadows caused by roadside greenery and buildings

The phenomenon of shadows is ubiquitous in the natural world and fundamentally contingent upon fluctuations in light. The formation of shadows is a consequence of the projection of light emitted by sources, such as the sun or lamps, which is then obstructed by objects. Consequently, shadows are cast in the direction of the light source’s projection toward the obstructing object (Al-Najdawi et al., 2012) (Figure 6). The presence of shadows can affect the perception of sensors, leading to visual changes that may obscure or blur the boundaries and obstacles on the road. This obscurity can, in turn, compromise the accuracy of vehicle positioning and path planning. Notably, during rapid environmental changes, shadows may considerably increase recognition errors within the system (Joshi and Papanikolopoulos, 2008). Currently, there are techniques for road shadow removal (Liu et al., 2021a). However, semantic segmentation models for shadow recognition still have deficiencies.

The effect of shadows on the accuracy of discerning traversable regions is highly pronounced and interferes with target detection in autonomous driving. In outdoor scenarios with sufficient light, the shadow cast by a vehicle becomes particularly prominent, which may result in the expansion and distortion of the vehicle’s target area. This distortion can result in the system failing to detect obstacles or incorrectly identifying them, thereby increasing safety risks (Guo et al., 2024). Furthermore, shadows affect distance perception in autonomous driving. The system may inaccurately perceive the distance to obstacles because of variations in light conditions, thereby compromising the vehicle’s obstacle avoidance capabilities. Pal et al. (2022) emphasized that under shadowing conditions, the error in distance estimation during autonomous driving considerably increases.

5.3.2 Road edge lines

The absence of delineating lines and the lack of a paved surface are defining characteristics of nonstructured roads (Figure 7), which typically lack a defined edge. In reality, such roads are common features of urban and rural landscapes (Alam et al., 2022). Further study is required to investigate the effects of nonstructured roads on the regional recognition capabilities of autonomous driving systems. The edge of the road is essential to the delineation of the vehicular driving area. Lane detection in driving scenarios serves as a crucial module for autonomous vehicles and advanced driver assistance systems. Recently, numerous studies have put forward sophisticated lane detection methods. However, most of these methods focus on detecting lanes from a single image and perform poorly when dealing with some extremely adverse conditions, such as heavy shadows and severely worn lane markings (Zou et al., 2019). For autonomous driving systems, accurately identifying the road edge is fundamental for planning a secure driving trajectory (Zhao et al., 2023). However, discrepancies in wear, contamination at the road edges, and variations in the reflective characteristics of different pavement materials can lead to errors in the driving systems’ assessments or omissions of certain road edge lines, thereby impeding their ability to delineate the driving area with precision. Abualsaud et al. (2021) proposed a lane detection method capable of detecting a variable number of lanes. Nevertheless, there are still research deficiencies in the segmentation of road edge lines.

The influence of road edge lines on the precision of detection within the driving area is initially manifested in autonomous driving path planning. Road edge lines can facilitate the smooth operation of autonomous vehicles. Accurate identification of the road edge is a prerequisite for the path planning of autonomous vehicles. If the system cannot identify the edge of the road accurately, the vehicle may deviate from the scheduled driving trajectory and even enter the nondriving area, potentially leading to unforeseen safety hazards (Babić et al., 2020). This scenario is reflected in the accuracy of autonomous driving recognition. Road edge lines are pivotal to the autonomous driving system. They delineate the boundaries of the viable area for the vehicle. Maddiralla and Subramanian (2024) investigated the effect of road edge detection on lane keeping in an autonomous driving system and found that by recognizing and understanding road edge lines, the system can accurately locate the vehicle position and determine a safe driving path. Tao (2016) analyzed the role of road margins in enhancing the navigation capabilities of autonomous vehicles. These lines indicate the area within which the vehicle should be driving and helps the system detect potential obstacles and other road users, thereby improving the navigation capabilities and traffic safety of the vehicle.

5.4 Result analysis

In the afternoon period when the sun is substantially inclined, shadow images were randomly captured at the intersection of Honghe Middle Road and Tu Shan Road in Chongqing City. The results are illustrated in Figure 8.

Three distinct image datasets from Chongqing were selected to evaluate road edge line conditions (missing, worn, or intact). Following manual segmentation of drivable areas, the results are presented in Figure 9.

Figure 10 displays the semantic segmentation results of road scenes generated by the model. We conducted statistical analyses to evaluate segmentation accuracy under different illumination conditions at the study intersection.

In general terms, the semantic segmentation results of the road under different lighting conditions were calculated with Formula 2. Given the different lighting conditions at the same intersection, under conditions with weak shadow influence, the average accuracy of semantic segmentation recognition was 95.9%. Under conditions with strong lighting and shadow influence, with the road creating distinct contrast between light and dark, the average accuracy of semantic segmentation recognition was 88.2%.

Specifically, shadow areas considerably affect the ability of the semantic segmentation model to recognize regions suitable for driving. The presence of shadows introduces complex local contrast changes in the image, and as a result, the model identifies redundant shadow boundary elements during the segmentation process (Figure 11). The superfluous boundary elements impede the model’s capability to accurately discern the actual road boundary, resulting in a blurring of the road boundary in the segmentation results. Hence, the overall accuracy of road recognition is adversely affected. Furthermore, this phenomenon increases the complexity of model processing and may also result in classification errors, thereby affecting the autonomous driving system’s precise understanding and decision-making capabilities regarding the road environment.

The mean accuracy of semantic segmentation recognition for images with clear road edge lines, which were found to influence the results, was 95.6%. By contrast, for images with missing or worn road edge lines, the mean accuracy of semantic segmentation recognition was 91.4%. Some result examples are shown in Figure 12.

Figure 13 reveals distinct impacts of road edge lines and shadows on drivable area recognition accuracy. Road edge line influence depends not only on line clarity but also on environmental color contrast. Notably, high-contrast regions maintain recognition accuracy even when edge lines are absent.

To facilitate a direct and comprehensive comparison of the experimental results under different built-environment conditions, Table 4 summarizes the quantitative results of drivable area detection accuracy reported in this section. As shown in Table 5, strong shadows lower the accuracy from 95.9% to 88.2% compared to weak shadow conditions. Missing or worn road edge lines lower the accuracy from 95.6% to 91.4% relative to clear edge lines. These aggregated data provide a clear overview of how each environmental factor affects segmentation performance.

TABLE 4. Ablation study results on the CamVid dataset (subtraction-based).

TABLE 5. Summary of drivable area detection accuracy under different built-environment conditions.

5.5 Recommendations for built-environment optimization

Based on our experimental results, we propose the following specific recommendations. The main optimization strategies are summarized for two key built-environment elements (road shadows and road edge lines), as well as for the color contrast between drivable and non-drivable areas. These recommendations aim to guide urban planners, road designers, and policymakers in retrofitting existing infrastructure or designing new roads that are more compatible with autonomous driving systems.

5.5.1 For road shadows (from trees or buildings)

Avoid planting tall trees or constructing buildings that cast persistent, sharp shadows onto drivable areas on roads with high autonomous vehicle traffic. If unavoidable, use textured pavement or reflective markings to maintain visual contrast within shadowed regions, and consider adjustable lighting systems such as dynamic streetlights to compensate for natural shadows during peak sunlight hours.

5.5.2 For road edge lines

Mandate high-durability, high-reflectivity materials for edge lines on roads intended for autonomous vehicle operation. Establish regular inspection and repainting cycles (e.g., every 6–12 months) for high-traffic autonomous driving corridors, and use raised pavement markers or retroreflective studs as auxiliary cues where paint wear is severe.

5.5.3 For color contrast (between drivable and non-drivable areas)

Design roads with a clear color difference between the asphalt (drivable) and shoulders/curbs (non-drivable), for example, dark grey asphalt versus light grey concrete or red-colored bus lanes. For existing low-contrast roads, apply colored surface treatments or high-contrast edge strips, and integrate color contrast requirements into future road design standards for autonomous driving.

6 Conclusions and discussion

6.1 1 Main conclusions

Accurate drivable area detection is critical for safe autonomous driving. This study investigates how built environment elements—particularly road shadows and edge lines—impact detection accuracy, and proposes an improved U-Net architecture with optimized expansion paths, refined pooling layers, and efficient convolutional kernels. Experimental results show the model achieves >90% average accuracy on the CamVid dataset. Application to our proprietary dataset reveals the following findings.

Road shadows significantly impact the semantic segmentation of drivable areas. The strong contrast between shadows and illuminated regions blurs boundaries, making accurate road boundary identification challenging and causing redundant information accumulation. The average recognition accuracy is 94.75% under minimal shadow influence, but decreases to 89.75% under strong lighting with substantial shadows. Therefore, future road designs for autonomous vehicles should minimize shadow-producing elements such as adjacent tall trees or buildings. To achieve this, we recommend avoiding planting tall trees or constructing buildings that cast persistent sharp shadows; where unavoidable, textured pavement or reflective markings and dynamic streetlights can help maintain visual contrast.

The clarity of road edge lines also considerably affects recognition accuracy. The average accuracy for images with clear edge lines is 95.6%, while it drops to 91.4% for missing or worn edge lines. The influence of edge lines depends not only on their clarity but also on the color contrast between the road and surrounding environment. In areas with strong contrast, the model can still accurately identify drivable areas even without discernible edge lines. However, under low contrast, accuracy declines significantly. Future road infrastructure should maintain clearly visible edge lines and sufficient color contrast. Practical measures include mandating high-durability, high-reflectivity materials for edge lines, using raised pavement markers or retroreflective studs where paint wear is severe, and designing roads with a clear color difference between asphalt and shoulders, with low-contrast roads treated by colored surface strips.

It should be noted that the proposed architectural improvements—dynamic pooling, hierarchical Dropout, and reduced initial channels—are generic modifications to the U-Net backbone rather than dataset-specific adjustments. Therefore, the improved network structure itself is transferable to other datasets (e.g., Cityscapes, BDD100K, or custom road scene datasets). However, the model weights (parameters) must be retrained on each new dataset, because semantic segmentation tasks involve different class distributions, image resolutions, and scene characteristics. The architecture does not require redesign; users can directly apply the same network structure and train from scratch or fine-tune on their target data. We provide our model implementation as a reusable framework to facilitate such transfer.

6.2 Limitations

This study has several limitations. First, the model’s recognition accuracy is susceptible to disruption when road patches are present on the road surface within the drivable area and when the color contrast between the patches and the road is substantial. This aspect requires further improvement. Second, one of the built environment elements selected in this study is the road edge line. However, road edge lines are influenced by other built environment elements, and research is needed to further understand the effect of missing or worn road edges on drivable area detection. Third, our self-collected dataset mainly covers daytime scenes with varying shadow intensities and does not include broader lighting conditions such as nighttime or twilight. The generalizability of our model under these scenarios requires future validation.

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.

Footnotes

Edited by: Jianhua He, University of Essex, United Kingdom

Reviewed by: Wenbin Zhu, Nanjing University of Science and Technology, China

Deivanathan R, VIT University Chennai, India