Visual Blind Spot Detection System for Bus Drivers Based on Lightweight Model

For a bus driver, the environment outside the bus door contains a serious visual blind spot. When a bus is parking, waiting for passengers boarding, and exiting the bus station, this blind spot requires the driver to maintain a high degree of attention. In order to improve the driver's perception of the blind spot in this area, a fisheye camera on the top of the bus door is used. A car-grade chip is selected to collect visual images and a lightweight deep learning model is deployed to identify the environment near the door, including objects such as pedestrians, bicycles and cars. Firstly, the collected YUV visual images are corrected for distortion, and then converted into RGB images. Then the RGB images are inputted into the trained deep learning model for human and vehicle recognition. The lightweight model is obtained after the lightweight optimization of the YOLO-FastestV2. And a good recognition accuracy is obtained. The neural network recognition of each image consumes about 60ms. In order to achieve the image collection and display rate of 25 fps, the frame skipping recognition method is adopted, which ensures the continuity of video display and a certain recognition accuracy. Finally, the performances of different deep learning models are compared so as to select a model suitable for the application scenarios and hardware resources. By providing the real time visual images with labels and an alarm signal when needed for the bus driver, the system could effectively improve the driver's ability of blind spot perception.