A Visual System for Blind Spot Detection and Warning for Mopeds

With the rapid development of computer vision, its related applications have been becoming more and more popular. One of the important reasons is that these applications can better assist people in various fields. In this article, the authors designed a novel blind spot detection system for mopeds. This article first described how to collect the data set containing cars and cycles, annotated the objects, and then converted it into YOLO (You Only Look Once) format. The authors then used the YOLOv5-Lite (You Only Look Once Version 5-Lite) model as the basic model for object detection. The authors used a cloud server to train the data set and obtained the optimal result, and then converted the optimal result into ONNX (Open Neural Network Exchange) format. After the authors implanted the ONNX format into the detection program, the program successfully realized the function of detecting specific objects in the blind spot areas for mopeds. The program also added a warning function, which will be activated when certain situations occur. Besides, the authors proposed a device called Blind Spot Detection and Warning Device (BSDWD), which is installed on mopeds and consists of two LED lights, a buzzer and a camera. After BSDWD is turned on, it will detect cars and cycles in the rear blind spot of the moped in real time. When the rear object enters the warning range, the warning system will be activated. Then the authors used BSDWD to conduct two sets of tests during the day and at night. The experimental results show that the accuracy of object detection is 71.7%. The real-time FPS (Frame Rate per second) of the camera is 6.37. Both sets of experiments can effectively detect objects in the rear blind spot areas, and the warning system can be activated successfully when the object enters the warning range.