Search for a command to run...
The growing incidence of forward collisions due to driver inattention highlights the critical importance of Advanced Driver Assistance Systems (ADAS) in enhancing vehicle safety through a robust Forward Collision Warning System (FCWS). Conventional techniques often rely on dual cameras or a combination of LiDAR, radar, and camera systems, increasing hardware cost and complexity. In contrast, this design uses a single camera for vehicle detection, tracking, and Time-to-Collision (TTC) calculation. A detailed comparative analysis revealed that the custom-trained You Only Look Once (YOLO) vehicle model optimized with TensorRT outperformed MobileNet Single Shot Multibox Detector (SSD), Region-based CNNs (R-CNN), ResNet-50, and ResNet-101, excelling in inference time, mean Average Precision (mAP), resolution, and initiation time. The model achieved an mAP of 65.8% with an inference time of 22 ms on the Berkeley DeepDrive (BDD 100K) dataset. Centroid-based tracking is employed to monitor detected vehicle centroids across consecutive real-time camera frames. The vehicle’s absolute speed, required for TTC calculation, is obtained via the OBD-II interface, enabling precise determination of the speed of the camera-equipped vehicle for accurate collision warnings. The results demonstrate that the low-cost single-camera system, enhanced by advanced optimizations and NVIDIA Jetson TX2 hardware, provides both robustness and reliability for FCWS applications.