A traffic sign recognition system is crucial for safely operating an autonomous driving car and efficiently managing road facilities. Recent studies on traffic sign recognition tasks show significant advances in terms of accuracy on several benchmarks. However, they lack performance evaluation in driving cars in diverse road environments. In this study, we develop a traffic sign recognition framework for a vehicle to evaluate and compare deep learning-based object detection and tracking models for practical validation. We collect a large-scale highway image set using a camera-installed vehicle for training models, and evaluate the model inference during a test drive in terms of accuracy and processing time. In addition, we propose a novel categorization method for urban road scenes with possible scenarios. The experimental results show that the YOLOv5 detector and strongSORT tracking model result in better performance than other models in terms of accuracy and processing time. Furthermore, we provide an extensive discussion on possible obstacles in traffic sign recognition tasks to facilitate future research through numerous experiments for each road condition.

In complex urban road conditions, stationary or illegally parked vehicles present a considerable risk to the overall traffic system. In safety-critical applications like autonomous driving, the detection of stopped vehicles is of utmost importance. Previous methods for detecting stopped vehicles have been designed for stationary viewpoints, such as security cameras, which consistently monitor fixed locations. However, these methods for detecting stopped vehicles based on stationary views cannot address blind spots and are not applicable from driving vehicles. To address these limitations, we propose a novel deep learning-based framework for detecting stopped vehicles in dynamic environments, particularly those recorded by dashcams. The proposed framework integrates a deep learning-based object detector and tracker, along with movement estimation using the dense optical flow method. We also introduced additional centerline detection and inter-vehicle distance measurement. The experimental results demonstrate that the proposed framework can effectively identify stopped vehicles under real-world road conditions.

We propose a lightweight cone-based Region of Interest (ROI) filtering method for camera–LiDAR fusion on distributed edge systems. Multiple Neural Processing Unit (NPU) nodes perform camera inference, and a central edge board combines their detections with LiDAR point clouds. The relative rotation is obtained from IMU quaternions under a common attitude reference and aligned sensor axes, while camera Field of View (FOV) parameters define the viewing rays. The method avoids a separate offline extrinsic-rotation estimation procedure, but it requires initial alignment, a measured translation vector, and timestamp-based synchronization. Because the rotation follows from the attitude streams rather than from a per-pair calibration session, a camera node can be added or re-aimed without a new calibration session, which lowers the setup cost of extending the system to further viewpoints. A cone membership test replaces four plane-normal dot products with a forward sign test and a squared angular cosine comparison that reuse the same axis–point dot product; on the same hardware, the mean per-camera ROI-filtering and clustering latency decreases from 6.02 to 4.46 ms, a 25.9% reduction. An adaptive threshold tightens the ROI boundary using the angular separation between neighboring detections. Across six overlap events in a parking scenario, pair-level separation succeeds in 2/6 cases (33.3%) with Pyramid and 5/6 cases (83.3%) with Cone+Adp. These preliminary results indicate improved ROI point selection for the tested configurations, rather than a general increase in intrinsic spatial separation capability.

Cargo-fall accidents caused by improper cargo loading have frequently occurred on highways in recent years, posing a growing threat to driver safety. Because enforcement typically relies on fixed inspection stations, some drivers illegally alter their cargo loads between checkpoints to evade overload inspections, and such changes are rarely secured properly, further increasing the risk of cargo-fall accidents. Current cargo inspection systems, however, are largely limited to labor-intensive manual checks or weight-based sensors that cannot distinguish between items of comparable weight. This paper presents an approach that detects cargo load changes on trucks by comparing depth maps acquired through a stereo vision system across consecutive passes. The proposed approach normalizes the raw depth map to correct for perspective distortion and aligns the current and previous cargo-bed regions through one-dimensional projection matching, enabling accurate comparison independent of cargo color or texture. We validate the approach through field experiments at SMTB under two test scenarios, one with uniform cargo items and one with diverse, irregularly placed items, and benchmark it against four widely used alignment algorithms: SIFT, ORB, AKAZE, and XFeat. The proposed algorithm reaches a combined detection rate of 91.49% (43/47), well above every conventional algorithm, at a processing speed nearly on par with ORB, the fastest of the compared methods. Notably, all four conventional algorithms consistently miss the same subset of events involving a tarpaulin, whose color and texture closely resemble the cargo bed, whereas the depth-based approach detects these events without difficulty. These findings offer practical guidance for building more reliable, real-world cargo load change detection systems.