Real‐Time Road Obstacle Detection System to Enhance Road Safety on African Roads

ABSTRACT

Globally, there has been a 5% decline in road accident fatalities. Integrating advanced technologies into vehicles in developed regions like Europe has significantly reduced road accident fatalities in these regions. This has played a pivotal role in reducing global road accident fatalities. However, the African road accident‐related fatalities have increased by 17%. Drivers' lack of sufficient technology to detect common African road obstacles is one of the leading causes of this increase in African road fatalities. These road accidents particularly affect the young and economically active population, impacting the continent's economic growth. Object detection models have effectively enhanced road safety in developed countries by detecting road obstacles. Unfortunately, these object detection models require substantial computational and memory resources, which limits their deployment on resource‐constrained edge devices. A real‐time road obstacle detection system is developed based on a YOLOv3 model in this study to address the rising accidents on African roads. The YOLOv3 model was trained on a custom dataset with African road‐specific obstacles. The trained model was deployed on an NVIDIA Jetson Nano for real‐world inference. The NVIDIA TensorRT half‐precision optimization was utilized to accelerate the model inference speed and reduce the model's memory usage while retaining the model's accuracy on the deployment platform. Experimental results reveal that deploying the model in TensorRT format reduced the inference time by 66%, achieving 68.8 ms (approximately 14.5 FPS, which meets the real‐time processing requirement for obstacle detection and collision warning systems), and the memory usage by 49.9% with a 0.35% drop in accuracy. The system offers an effective and cost‐effective solution on affordable hardware to improve road safety across African roads.