Adaptive calibration and spatio-temporal filtering for robust highway wrong-way driving detection

1 Introduction

Wrong-way driving detection is not only a computer vision task but also a key component of proactive traffic safety management, helping shift highway operations from post-incident response to early warning and prevention. Camera-based highway surveillance is attractive for this purpose because it can reuse existing monitoring infrastructure and support large-scale deployment without modifying the pavement or requiring additional in-vehicle equipment. However, operational roadside scenes remain challenging: camera viewpoints may rotate after wind, vibration, or maintenance; vehicles may be partially occluded by trucks; and stationary vehicles with flashing lights may induce motion-like visual changes.

Early wrong-way driving detection systems relied mainly on physical sensors or low-level motion cues. Sensor-based approaches used radar data analysis Shi and Abdel-Aty (2015), high-resolution inductive loop detectors Xu et al. (2016), or real-time traffic monitoring data Chakraborty et al. (2017). Other systems used pressure-sensitive strips to infer driving direction from the order of hardware triggers Chen et al. (2023). Although these solutions can be effective in controlled deployments, they require dedicated roadside hardware and may be affected by pavement condition, weather, and maintenance costs. Computer vision methods based on optical flow, road disparity maps, and MedianFlow tracking have also been studied Monteiro et al. (2007); Ha et al. (2014); Jain and Tiwari (2015); Şentaş et al. (2019). These methods reduce hardware dependence but remain sensitive to illumination changes, shadows, tunnel-light flicker, and pixel-level noise, which can be particularly problematic for real-time processing at roadside edge nodes.

Recent deep learning methods have improved the accuracy of vehicle detection and tracking Liu et al. (2025); Zeng et al. (2026). Two-stage detectors, single-stage detectors such as single-shot detectors, and the YOLO series have enabled more practical vehicle localization pipelines Girshick et al. (2014); Girshick (2015); Ren et al. (2016); Liu et al. (2016). In wrong-way driving analysis, object detectors are often cascaded with multi-object tracking methods such as DeepSORT, Kalman filtering, lane-partitioning modules, road-boundary constraints, and attention-enhanced detectors Zuraimi and Zaman (2021); Hanzla et al. (2024); Haghighat and Sharma (2023); Suttiponpisarn et al. (2022); Liu and Shen (2025); Rahman et al. (2020); Mampilayil and Rahamathullah (2019). These developments have strengthened trajectory-based behavior recognition, but many systems still assume a stable camera viewpoint and infer wrong-way motion primarily from centroid displacement or bounding-box geometry.

Two deployment-oriented gaps remain insufficiently addressed. First, highway surveillance cameras are not always geometrically stable. A small physical rotation can invalidate hard-coded lane regions and reference directions, causing normally moving vehicles to be misclassified as wrong-way targets. Second, trajectory-based decisions can confuse true wrong-way motion with pseudo-motion produced by short-term occlusion, bounding-box drift, or local light flicker from stationary vehicles Liu et al. (2023), Liu et al. (2018). These two issues directly affect the operational reliability of roadside intelligent infrastructure and reduce the trustworthiness of automated traffic safety alarms.

This study addresses these gaps by proposing an adaptive calibration and spatio-temporal cascaded filtering framework. The main contributions are as follows. First, the framework constructs lane-level driving-direction references by combining road semantic segmentation with historical vehicle trajectories, reducing dependence on manual lane-direction annotation. Second, it introduces a lightweight online camera-rotation perception mechanism based on difference hashing, enabling automatic recalibration after viewpoint changes. Third, it designs a cascaded verification strategy that combines temporal continuity filtering with perceptual feature-difference analysis to suppress false alarms caused by tracking jitter and stationary vehicles. Finally, a dedicated highway surveillance dataset covering highways, tunnels, and camera-rotation scenarios is constructed to evaluate the method under practical operational conditions.

2 Materials and methods

2.1 Dataset construction

To address the limited availability of highway wrong-way driving samples and the lack of environmental diversity in public datasets, this study constructed two highway surveillance datasets for method development and evaluation: a wrong-way driving detection dataset and a Feature Difference Mean dataset. The wrong-way driving detection dataset contains 99 video sequences collected from operational highway surveillance cameras and covers three challenging scene categories: open highways (60 videos), tunnels (17 videos), and camera-rotation events (22 videos). To strictly prevent data leakage and guarantee independent evaluation, the 99 video sequences were partitioned at the video sequence level (rather than frame-level random sampling) into three disjoint subsets: a Training set (51 videos), a Validation set (20 videos), and an Independent Test set (28 videos comprising over 5 h of surveillance footage and 43 independent wrong-way events defined by temporal intervals ≥ 15 seconds). The video-level distribution across scene categories is maintained proportionally across all splits, ensuring that frames from the same video sequence never appear in both the tuning and testing sets.

Figure 1 shows representative examples, and Table 1 summarizes the scene distribution. The videos cover daytime, nighttime, tunnel, occlusion, stationary-vehicle, and camera-viewpoint-change scenarios. They were recorded at resolutions of 1920 × 1080 and 3840 × 2160 with an original frame rate of 25 fps. To reduce storage and computation while emulating low-frame-rate roadside deployment, one frame was sampled every 12 frames, resulting in an effective frame rate of approximately 2.08 fps. The Feature Difference Mean dataset was constructed from the same intelligent highway surveillance environment and includes stationary and moving vehicles for training and evaluating the secondary verification module.

TABLE 1. Scene composition of the highway wrong-way driving detection dataset.

Bold values indicates the best performance/the proposed methods.

As shown in Table 2, the Feature Difference Mean dataset contains 20,864 training samples, 1,854 validation samples, and 465 testing samples. This distribution supports parameter selection and performance evaluation for the secondary verification module.

TABLE 2. Composition of the feature difference mean dataset.

Table 3 presents illustrative examples of the V mean feature-difference values for stationary and moving vehicles. The verification pipeline operates on the [0, 255] intensity scale described below; only the values displayed in Table 3 are rescaled for readability as V mean disp = V mean / 255, yielding the displayed [0, 1] range. The wrong-way driving videos were annotated at the vehicle-track level. A detection is considered a true positive when the reported wrong-way vehicle ID matches a manually annotated wrong-way vehicle ID. A false positive is counted when an unannotated or normally moving vehicle track is reported as wrong-way driving, and a false negative is counted when an annotated wrong-way vehicle is missed. This track-level protocol reflects the operational goal of roadside alarm generation, where each detected wrong-way vehicle should correspond to a real traffic safety event rather than to an isolated frame-level response.

TABLE 3. Examples of display-normalized feature difference mean values for stationary and moving vehicles ( V mean disp = V mean / 255 ).

2.2 Proposed framework

As illustrated in Figure 2, the proposed framework contains two coupled modules: online dynamic calibration and cascaded filtering detection. The online dynamic calibration module continuously monitors the video stream, detects camera perspective rotation, and updates lane-level driving-direction references when the surveillance viewpoint changes. The cascaded filtering detection module then verifies candidate wrong-way trajectories through a process that integrates temporal continuity filtering, perceptual difference map analysis, and max-margin adaptive boundary matching. This design targets both viewpoint-induced reference errors and object-level pseudo-motion, improving robustness under realistic highway surveillance disturbances.

To bridge the gap between ideal computer vision techniques and real-world roadside infrastructure reliability, the primary strength of this framework lies in its system-level architecture design that effectively balances the adaptation of established perception models with custom-designed engineering logic.

• Adapted Baseline Perception Techniques: We adapt standard, highly effective computer vision backbones to establish base visual perception, including DeepLabv3+ for road semantic segmentation, YOLOv8 coupled with ByteTrack for robust multi-vehicle detection and tracking, and an ImageNet-pretrained VGG-16 network for localized deep feature extraction.
• Custom ITS System-Level Modules: Built upon these adapted backbones, the main methodological contribution lies in their integration with two deployment-oriented mechanisms: (1) an online dynamic calibration mechanism driven by difference hashing (DHash) to perceive substantial viewpoint changes and update lane direction references ( D ⃗ l ) without manual re-annotation, and (2) a spatio-temporal cascaded filtering strategy that combines temporal persistence filtering with ROI feature-difference ( V mean ) max-margin classification to reject pseudo-motion false alarms.

2.3 Online dynamic calibration

As illustrated in Figure 3, the Online Dynamic Calibration module consists of four sequential components to achieve adaptive baseline calibration: lane partitioning, lane direction determination, camera rotation detection, and dynamic lane direction adjustment. The overall workflow begins with an initialization phase, where the module performs lane partitioning to establish spatial lane regions and executes lane direction determination to mine reference driving directions from historical vehicle trajectories. During routine operation, the camera rotation detection component continuously monitors perspective variations online. Once a perspective change is captured, the dynamic lane direction adjustment strategy immediately rectifies the driving direction references to maintain spatial consistency.

2.3.1 Lane partitioning and direction determination

For lane partitioning, the framework first uses DeepLabv3+ Chen et al. (2018) to extract drivable road areas and generate a binary road mask, which is then refined by morphological closing and opening operations. Based on this mask, Union-Find connected-component labeling is used for multi-lane clustering. Finally, following a “left-first priority” criterion, the clustered lanes are sorted in ascending order according to their horizontal centroid coordinates and assigned unique topological indices from 1 to N, as illustrated in Figure 4.

For lane direction determination, the framework extracts continuous vehicle trajectories by integrating YOLOv8 object detection with ByteTrack multi-object tracking Zhang et al. (2022). YOLOv8 provides vehicle localization, while ByteTrack maintains trajectory continuity under dynamic occlusion. To reduce interference from temporary lane changes or short-term heading deviations, the framework aggregates historical vehicle displacement vectors v ⃗ i within lane l to calculate the standardized driving direction vector D ⃗ l:where v ⃗ i represents the i-th historical vehicle motion displacement vector within lane l, N l denotes the total number of valid displacement vectors that have historically traversed lane l, and D ⃗ l is the compiled standard driving direction vector for that lane.

D ⃗ l = 1 N l ∑ i = 1 N l v ⃗ i ‖ v ⃗ i ‖

Note that D ⃗ l is established purely from historical trajectory aggregates to prevent reference contamination, rather than relying on the current candidate target’s instantaneous motion vector v ⃗ t. After D ⃗ l has been compiled, its longitudinal component D l , y is used only to assign a semantic direction label to lane l: D l , y < 0 denotes the approaching direction (moving closer to the camera perspective), whereas D l , y > 0 denotes the receding direction. In contrast, v ⃗ t = ( v t , x , v t , y ) is reserved for the instantaneous displacement of an individual candidate vehicle at frame t and is used later for wrong-way decision making through its relation to the fixed lane reference D ⃗ l.

2.3.2 Camera rotation perception

To support lightweight real-time monitoring, the framework uses Difference Hashing (DHash) Di et al. (2022) to detect camera rotation. In implementation, each frame is converted to grayscale and resized to 9 × 8, producing a 64-bit hash code. The DHash workflow contains two stages: image preprocessing and inter-frame similarity calculation. During preprocessing, the current video frame is converted to grayscale and downsampled to an M × N low-frequency representation. The framework then computes horizontal grayscale gradients between adjacent pixels. By comparing the grayscale values of neighboring pixels, it generates a B-bit binary hash code H f according to the following gradient encoding rule:where G ( x , y ) denotes the grayscale value of the low-frequency image at spatial coordinates ( x , y ), H f ( x , y ) represents the corresponding encoded bit within the profile, M and N denote the target width and height dimensions of the downsampled image, respectively, and B = ( M − 1 ) × N represents the total compiled bit-length of the hash code. In the inter-frame similarity calculation stage, to mitigate sporadic single-frame interferences triggered by environmental illumination variations, instantaneous headlight flickering, or environmental noise, the framework evaluates temporal structural variations by computing the average Hamming distance a v g d:where H f − 1 and H f − 2 represent the B-bit binary hash codes of the previous frame and the frame prior to that, respectively, while Hamming ( ⋅ ) denotes the standard Hamming distance operation. When the computed average Hamming distance satisfies a v g _ d > σ d (where σ d is the configured experimental determination threshold), the framework determines that a physical perspective rotation has occurred in the surveillance camera. To prevent the framework from being repeatedly triggered due to the unstable transition states of consecutive frames during the camera rotation transition period, a detection dormancy lock mechanism is automatically activated immediately after successfully confirming the rotation behavior.

H f x , y = 1 , if G x , y > G x + 1 , y 0 , otherwise

a v g _ d = Hamming H f , H f − 1 + Hamming H f , H f − 2 2

2.3.3 Dynamic lane direction adjustment

Once a camera perspective rotation is confirmed, the framework executes the dynamic lane direction adjustment strategy to rectify the direction references. The current strategy focuses on large viewpoint reversal or rotation events that change the incoming/outgoing relationship of the lane direction. Small camera jitter, gradual viewpoint drift, and zoom changes are treated as environmental perturbations rather than as full direction-reversal events.

First, the DeepLabv3+ network is reinvoked on the post-rotation video stream to reconstruct the updated lane spatial partitioning. The direction reference is then updated according to the number and relative positions of the detected road regions. For a single-direction highway scene, the correct driving direction is inverted after a direction-changing camera rotation:where D s 0 and D s 1 denote the single-roadway direction before and after camera rotation, respectively. For a two-roadway scene, the spatial left/right relationship in the camera view is re-established after lane partitioning, and the left and right roadway references are updated aswhere D p l 0 and D p r 0 denote the pre-rotation left and right roadway direction references, and D p l 1 and D p r 1 denote their post-rotation counterparts after the updated lane partitioning. Finally, a temporal locking mechanism suspends motion-vector updates for a configured number of frames to ensure tracking baseline stabilization before passing the newly calibrated spatial benchmarks to the Cascaded Filtering Detection module.

D s 1 = − D s 0 ,

D p l 1 = D p l 0 , D p r 1 = D p r 0 ,

2.4 Cascaded filtering detection

As illustrated in Figure 5, the Cascaded Filtering Detection module executes a progressive verification workflow to enhance the framework’s anti-interference capability against spatial-geometric variations. First, the temporal continuity filtering stage utilizes the persistence characteristics of temporal trajectories to filter out instantaneous tracking jitter and sporadic bounding box noise. Subsequently, the perceptual difference map analysis and max-margin adaptive boundary matching stages perform fine-grained object-level analysis on the candidate targets to eliminate complex stationary false alarms. The detailed implementations of these validation stages are elaborated in the following subsections.

2.4.1 Temporal continuity filtering

For each target tracking trajectory i maintained by the multi-object tracking module, the framework independently maintains a temporal wrong-way status counter n x ( i ). During the routine vehicle tracking workflow, the wrong-way determination operator D is formally defined as the inner product between the current motion displacement vector v ⃗ t of the target and the standardized reference direction vector D ⃗ l of its corresponding lane l:where v ⃗ t denotes the motion displacement vector of the tracking target at the current frame t, and D ⃗ l represents the standardized reference driving direction vector compiled for lane l. On this basis, dynamic updates of the temporal wrong-way status counter n x ( i ) are strictly executed based on the positive or negative sign of the inner product result, according to the following formulation:

D = v ⃗ t ⋅ D ⃗ l

n x i = n x i + 1 , if D < 0 0 , otherwise

An inner product result of D < 0 indicates that the instantaneous motion direction of the target is opposite to the standard lane direction; hence, the single-frame driving status of this target is flagged as a suspected wrong-way anomaly, triggering an increment in the counter. Conversely, if D ≥ 0, it signifies that the target is driving normally within its designated lane bounds, and the framework immediately resets the counter to zero. Only when this counter continuously grows with the progression of video frames and satisfies n x ( i ) ≥ t h r (where t h r represents the configured temporal persistence threshold) will the target be formally transitioned into the suspected wrong-way candidate set and delivered to the next stage for deep semantic verification. Through this progressive tracking constraint, the framework can effectively isolate isolated single-frame bounding box jitter and sporadic tracking noise.

2.4.2 Perceptual difference map analysis

Aiming at the suspected wrong-way candidate targets output from the primary filtering stage, this section introduces an object-level deep semantic feature difference perception analysis mechanism to effectively suppress pseudo-wrong-way false alarms. The processing pipeline directly utilizes the target trajectories obtained from the previous tracking stage. Specifically, for any suspected wrong-way trajectory, the framework captures its corresponding start frame I start and end frame I end as bounded by the primary filtering phase. Following semantic feature comparison and perceptual feature-loss formulations Di Biase et al. (2021), the deep semantic feature difference L feat ( I start , I end ) between these two temporal anchors is computed via the VGG-16 network pre-trained on ImageNet Russakovsky et al. (2015), which is mathematically defined as follows:where ϕ k represents the feature map of the k-th layer of the VGG-16 network, M k denotes the total number of elements in the k-th layer, and N is the total number of operational layers utilized. The computed feature difference tensor is subsequently normalized to the [ 0,255 ] grayscale range to yield a global spatio-temporal feature difference map.

L feat I start , I end = ∑ k = 1 N 1 M k ‖ ϕ k I start − ϕ k I end ‖ 1

Subsequently, the framework acquires the precise target bounding box position at the start frame I start to crop the corresponding local ROI feature difference map, denoted as T diff. To quantitatively evaluate whether a significant physical displacement exists during this time interval, the framework computes the arithmetic mean value V mean of all feature pixel elements within this localized ROI:where T diff ( j ) represents the feature difference value at the j-th pixel in the cropped local difference map, and N pixel is the total number of feature pixels contained within the target bounding box at the start frame. Finally, the quantified V mean is evaluated against the classification decision hyperplane to judge whether the target exhibits a significant movement, thereby accurately distinguishing true moving wrong-way vehicles from stationary interferences.

V mean = 1 N pixel ∑ j = 1 N pixel T diff j

2.4.3 Max-margin adaptive boundary matching

To construct a robust classification boundary to block the interference of stationary false-positive targets, this section introduces a max-margin adaptive strategy based on the quantified motion intensity. The input feature x i for the i-th candidate target bounding box is directly defined as its local feature difference mean V mean obtained from the secondary verification stage:

x i = V mean i

The corresponding category label y i indicates whether the candidate target represents a true wrong-way driving or a stationary interference, which is formulated as follows:

y i = 1 , true wrong - way driving − 1 , stationary interference

The max-margin criterion is designed to separate the positive and negative samples as far as possible by solving the following primal optimization problem:where M represents the total number of candidate target boxes, ω and b are parameters used to define the decision hyperplane, and λ is the regularization coefficient used to adjust the regularization strength. To further exploit the non-linearity in discrimination, the optimization problem is transformed into its equivalent dual form, and a non-linear kernel function is introduced to measure feature similarity: where α is a Lagrange multiplier vector, e denotes a vector of all ones, C is the constant penalty parameter, and Q is a positive semi-definite matrix whose elements are calculated as follows: where γ is the parameter determining the width of the kernel function.

min ω , b ∑ i = 1 M max 0,1 − y i ω T x i + b + λ ‖ ω ‖ 2 2

min α 1 2 α T Q α − e T α

s.t. y T α = 0 , 0 ≤ α i ≤ C , i = 1,2 , … , M

Q i j = y i y j K x i , x j

K x i , x j = exp − γ ‖ x i − x j ‖ 2

This linearly constrained convex quadratic programming problem can be successfully solved by a standard optimization software package. Through offline training on the constructed feature-difference dataset, the solver converges to an optimal classification hyperplane. If the final decision output of the secondary verification classifier is 1, the framework identifies the target as a true wrong-way driving event, outputs the final detection bounding box, and triggers an active warning signal. If the output is −1, the target is treated as a stationary vehicle or tracking drift interference and is eliminated.

3 Results

3.1 Experimental setup

The framework was implemented in Python 3.8 with PyTorch 1.12.1, CUDA 11.3, and CUDNN 8.9.1 on Windows 10. Experiments were conducted on a workstation equipped with an AMD Ryzen 7 3800X eight-core CPU, an NVIDIA GeForce RTX 2080 SUPER GPU with 8 GB memory, and 64 GB RAM.

To characterize computational efficiency, detailed runtime breakdowns and memory footprints were measured under FP16 precision on the workstation described above. The complete pipeline achieves an average processing throughput of 48.5 FPS (20.6 ms/frame, incorporating video decoding and Host-to-Device memory transfer overheads). Per-module processing latencies are as follows: video frame decoding and pre-processing require 3.2 ms/frame, YOLOv8 object detection takes 9.8 ms/frame, ByteTrack multi-object tracking requires 2.6 ms/frame, and the DHash rotation perception module incurs <0.1 ms/frame. The VGG-16 feature-difference verification module takes 12.4 ms per invocation; because it is triggered on demand only for candidate wrong-way trajectories, its amortized per-frame overhead remains around 5.0 ms. DeepLabv3+ road re-segmentation operates as an event-driven task requiring 142.5 ms only upon confirmed camera rotation events. The peak runtime memory footprint is 3.8 GB GPU VRAM (including the CUDA context, pre-allocated memory pools, model weights, and dynamic feature buffers) and 3.4 GB CPU RAM. The 48.5-FPS value represents computational throughput on the evaluated workstation rather than a direct benchmark on a roadside edge device; the detection experiments use the approximately 2.08-FPS sampled surveillance stream described in Section 2.1. Thus, these measurements demonstrate real-time computational feasibility on the evaluated platform and provide a reference for subsequent edge-device optimization.

For parameter configuration, the average Hamming distance threshold for camera-rotation detection is set to 25.5, and the detection dormancy lock period is set to 5 frames to avoid repeatedly triggering on a single camera-rotation transition.

The consecutive wrong-way frame threshold t h r ∈ [ 1,8 ] was evaluated on the Validation set to balance alarm latency and false-alarm suppression (Table 4). As t h r increases, transient tracking noise is filtered out, with t h r = 5 achieving the optimal balance and the highest F1-score of 96.77%. Consequently, t h r = 5 was fixed for all independent test set evaluations.

TABLE 4. Parameter sensitivity under different temporal threshold values on the validation set.

Bold values indicates the best performance/the proposed methods.

The framework is evaluated using event-level precision, recall, and F1-score. A detection is counted as a true positive when the detected wrong-way driving event ID matches the annotated ID. If a video contains an unannotated vehicle ID reported as wrong-way driving, the result is counted as a false positive. If an annotated wrong-way vehicle ID is not detected, the result is counted as a false negative. This definition is consistent with practical traffic management, where the system is expected to trigger one reliable alarm for each true wrong-way vehicle track.

The wrong-way dataset is annotated at the vehicle-track/event level for operational alarm evaluation rather than as a frame-level detection and identity benchmark. Accordingly, detector-only mAP and tracker-only HOTA/IDF1 are not reported in the present study. Instead, the influence of alternative tracking pipelines is evaluated through end-to-end YOLOv8+DeepSORT and YOLOv8+ByteTrack baselines under the same wrong-way decision setting, while the proposed dynamic-calibration and cascaded-filtering modules are examined through the component-oriented comparisons and ablations below. This evaluation scope is stated explicitly to avoid interpreting the reported event-level metrics as standalone detector or tracker benchmarks.

3.2 Experimental analysis

3.2.1 Wrong-way driving detection comparison

To evaluate the wrong-way detection performance, the proposed framework was compared against representative state-of-the-art baselines on the independent Test set, including classical centroid-based approaches (RT-TVD Şentaş et al. (2019), YOLO + ROI-Tracking Rahman et al. (2020), Y5+Kalman-Filter Vardhan et al. (2023)) and widely-adopted modern multi-object tracking pipelines (YOLOv8+DeepSORT and YOLOv8+ByteTrack). Specifically, in these two modern tracking baselines, wrong-way motion is determined by calculating the cosine similarity (or inner product) between the vehicle’s short-term trajectory displacement vector—derived from Kalman-filtered bounding-box centroids across consecutive frames—and a pre-configured, static lane direction vector. The quantitative results are summarized in Table 5.

TABLE 5. Comparison with existing wrong-way driving detection methods.

Bold values indicates the best performance/the proposed methods.

As shown in Table 5, the proposed framework achieves an overall Precision of 97.62%, Recall of 95.35%, and an operational false alarm rate of 0.12 FAR/h across 28 long-duration test videos (>5 h). The main advantage is most evident under camera rotation, where methods relying on fixed direction references experience substantial performance degradation and increased false alarms. In this subset, the proposed dynamic-calibration mechanism maintains 93.33% recall with 0.15 FAR/h, whereas the compared static-reference tracking baselines show markedly higher false-alarm rates. In tunnel scenes, the proposed framework provides accuracy comparable to the strongest tracking baselines while reducing FAR/h, highlighting that the intended benefit of the cascaded verification stage is operational false-alarm suppression rather than uniformly maximizing every accuracy metric in every sub-scenario. Overall, the results support the value of combining adaptive reference calibration with secondary verification under heterogeneous highway surveillance conditions.

The proposed framework achieves an average alarm latency ( T alarm ) of 9.8s and a processing throughput of 48.5 FPS on the evaluated workstation. In highway scenes, the YOLOv8+ByteTrack baseline triggers alarms slightly faster (7.8s vs. 9.2s), whereas the proposed framework introduces secondary feature-difference verification to reduce false alarms. Under tunnel lighting and camera-rotation conditions, the proposed framework reports alarm latencies of 9.8s and 10.2s, respectively, while retaining substantially lower FAR/h in the rotation subset. These results indicate a practical trade-off between alarm delay and false-alarm suppression. Because the throughput measurements were obtained on an RTX 2080 SUPER workstation, they are interpreted as evidence of real-time computational feasibility on the evaluated platform rather than as a direct edge-device deployment benchmark.

Figure 6 provides qualitative examples of these differences. RT-TVD Şentaş et al. (2019) captures the wrong-way driving event in one sequence but produces false alarms when tracking becomes unstable. YOLO + ROI-Tracking Rahman et al. (2020) correctly detects the target in one case but is affected by bounding-box drift when large vehicles occlude stationary vehicles. The Y5+Kalman-Filter baseline Vardhan et al. (2023) is affected by illumination variation and static-target jitter, producing false positives in low-light scenes. In contrast, the proposed framework suppresses false alarms induced by stationary vehicles, local light changes, and tracking instability, which is critical for practical highway surveillance deployments.

To validate sub-module effectiveness, the Online Dynamic Calibration (DC) and Cascaded Filtering Detection (CFD) modules were evaluated on the 28-video Independent Test set (Table 6). On Non-Rotation scenes, the standalone CFD module achieved 96.43% precision and recall (27/28) with a 0.11 FAR/h, proving its capability in suppressing pseudo-motion false alarms. On Camera-Rotation scenes, activating the DC module restored directional alignment, raising recall from 46.67% (7/15) to 93.33% (14/15) and precision from 63.64% (7/11) to 93.33% (14/15), while reducing false alarms from 3.45 to 0.15 FAR/h. Integrated across all test sequences, the framework achieved an overall Precision of 97.62% (41/42), Recall of 95.35% (41/43), F1-score of 96.47%, and 0.12 FAR/h.

TABLE 6. Sub-module performance evaluation across scene sub-categories.

Bold values indicates the best performance/the proposed methods.

3.2.2 Camera rotation detection comparison

To address the misalignment of lane driving direction references caused by camera rotations, this study conducts a comparative analysis of different image similarity evaluation algorithms. The comparative scheme is designed using the control variable method, wherein only the camera-rotation perception strategy within the lane-direction dynamic calibration module is substituted, while the logic and parameter settings of all other sub-modules remain identical.

Table 7 reports the end-to-end wrong-way driving performance obtained when the camera-rotation perception strategy is replaced while the remaining pipeline is kept unchanged. This comparison measures the downstream effect of each rotation-perception choice on the complete system rather than standalone camera-motion estimation accuracy. DHash provides the best overall F1-score in this end-to-end setting. The design motivation is that the calibration module requires a lightweight trigger indicating that a substantial viewpoint change has occurred; it does not require dense motion fields or explicit geometric transform estimation at every frame. The qualitative design comparison in Table 8 further explains this task-specific choice.

TABLE 7. Comparison of camera-rotation perception strategies.

Bold values indicates the best performance/the proposed methods.

TABLE 8. Design-oriented comparison of camera-viewpoint change analysis techniques.

3.2.3 Ablation study

To evaluate the contribution of each component, this study conducts an ablation analysis on the highway video dataset. The evaluated components include online dynamic calibration (DC), temporal continuity filtering (TF), perception difference map analysis (DA), and max-margin adaptive boundary matching (BM). The quantitative results are summarized in Table 9. The complete architecture (DC + TF + DA + BM) achieves the highest F1-score of 96.47%, while removing any component degrades performance. The DC module maintains the validity of lane-direction references after camera rotation. The TF module removes short-lived motion noise before verification is performed. The DA and BM modules then reject stationary-vehicle and bounding-box-drift false positives through object-level feature-difference analysis and adaptive boundary matching. Overall, calibration reduces reference errors, temporal filtering reduces transient noise, and feature-difference boundary matching reduces object-level pseudo-motion.

TABLE 9. Ablation results of core framework components.

Bold values indicates the best performance/the proposed methods.

4 Discussion

4.1 Relevance to future transportation systems

The results show that wrong-way driving detection for highway surveillance should be treated not only as a visual recognition problem, but also as an operational reliability problem for future transportation infrastructure. In practical deployments, false alarms can reduce operator trust and delay real responses, whereas missed detections can reduce the value of early-warning systems. The proposed framework improves reliability by explicitly addressing two deployment disturbances that are common in roadside monitoring: camera viewpoint changes and pseudo-motion induced by tracking instability or stationary vehicles. This deployment-oriented design makes the framework relevant to roadside intelligent infrastructure, where automated warnings must be accurate, timely, and robust enough to support proactive traffic management rather than merely post-event review.

4.2 Effect of cascaded filtering

The cascaded filtering module reduces false alarms by separating candidate generation from candidate verification. Temporal continuity filtering first removes isolated frame-level jitter, while feature-difference perception evaluates whether the target exhibits meaningful object-level motion. This is particularly effective for stationary vehicles affected by bounding-box drift, occlusion, or flashing lights. The ablation study confirms that the best performance is obtained only when calibration, temporal filtering, and perceptual verification are used together, demonstrating that the modules solve complementary sources of error.

4.3 Analysis of DHash selection

The DHash module is used as a lightweight structural-change trigger rather than as a replacement for general-purpose camera-motion estimation. Optical flow, feature matching, and homography estimation can provide richer motion or geometric information, including motion fields, correspondences, or explicit transforms. However, the present calibration logic only requires an inexpensive indication that a sufficiently large viewpoint change has occurred so that lane segmentation and direction references can be refreshed. DHash addresses this narrower requirement by encoding spatial grayscale gradients into a compact 64-bit hash and comparing consecutive frames through Hamming distance.

Table 8 therefore provides a design-oriented qualitative comparison of the information returned by these alternatives and their relative per-frame computational requirements; it should not be interpreted as a standalone accuracy benchmark of optical flow, feature matching, or homography estimation. In the proposed implementation, DHash is coupled with a Hamming-distance threshold of 25.5 and a 5-frame dormancy lock, and the end-to-end effect of alternative rotation-perception strategies is reported separately in Table 7. This task-matched design avoids continuously estimating a dense motion field or geometric transform when the downstream module only needs a recalibration trigger. Its limitation is that DHash does not estimate the magnitude or geometric parameters of camera motion and is therefore most suitable for the substantial viewpoint-change events considered in this study rather than gradual drift, zoom, or fine geometric stabilization.

4.4 Limitations

Several limitations remain. First, the current lane-direction calibration strategy mainly considers single-roadway and two-roadway highway scenes, which are common in the constructed dataset but may not cover complex interchanges, ramps, or urban intersections. Second, the camera-rotation module is designed for large viewpoint changes that alter the incoming/outgoing direction relationship; gradual drift, zooming, and partial road-area occlusion require further investigation. Third, the robustness of the feature-difference verification module may decrease under heavy rain, dense fog, strong road reflections, or severe image degradation, because these conditions can shift the distribution of V mean. Fourth, a single-camera view cannot fully resolve long-term occlusion, high-curvature road sections, or spatial blind spots caused by large vehicles. Fifth, because the present dataset was annotated for vehicle-track/event-level wrong-way alarms rather than for frame-level detector and identity-tracking benchmarks, standalone detector mAP and tracker HOTA/IDF1 are not reported; dedicated component-level annotations would strengthen future analysis. Finally, the current evaluation focuses on highway surveillance videos from the constructed dataset; broader validation across regions, camera types, weather conditions, and embedded edge hardware would further strengthen generalizability and deployment evidence.

5 Conclusion

This study proposed an adaptive calibration and spatio-temporal filtering framework for robust highway wrong-way driving detection. The framework constructs lane-level reference directions from road semantic segmentation and historical trajectories, detects camera rotation through lightweight difference hashing, and verifies suspected wrong-way targets through a cascaded strategy integrating temporal continuity filtering, perceptual difference map analysis, and max-margin adaptive boundary matching. Experiments on a dedicated 99-video highway surveillance dataset demonstrate that the proposed framework achieved 97.62% precision, 95.35% recall, and a 96.47% F1-score, while also reducing false alarms under camera rotation, occlusion, tracking jitter, and stationary-vehicle interference. Future work will investigate weather-aware feature normalization, dedicated detector/tracker component benchmarking, multi-camera cooperative detection, cross-region adaptation, and edge-device optimization to support more reliable deployment in large-scale intelligent transportation systems.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. For language check and polishing.

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: Konstantinos Koufos, Queen’s University Belfast, United Kingdom

Reviewed by: Agha Asim Husain, ITS Engineering College, India

Hejun Lv, Beihang University Institute of Unmanned System, China