Deep Learning-based Road Object Detection for Collision Avoidance in Autonomous Driving

With the population in metro cities rising, traffic congestion has become a significant issue, resulting in an alarming number of traffic collisions and road accidents. Given the advancements in automated vehicles, it is crucial to have a vehicle detection system that is equipped to handle the various challenges that arise on the road, particularly when it comes to traffic collisions. Although many datasets are available to support object detection for traffic monitoring and management, the suitability of these datasets for specific weather conditions across the globe needs to be analyzed. We present Canadian vehicle datasets (CVD) and analyze the YOLOv8-based deep learning model using CVD. Thales, Canada, captured street-level videos in Quebec City, Canada, with a vehicle equipped with a high-quality RGB camera. The videos are taken day and night in all four seasons: rain, snow, hazy, and bright sunlight. We created 27378 labels for 11 classes for 8000 images. The YOLOv8 model trained on the publicly available roboFlow dataset was compared to the model trained on the combined roboflow and our weather-specific CVD.