Alarm Algorithm for vehicle blind zone based on lightweight network

Abstract:
The a-pillar of a car plays the role of protecting the whole inner structure of the car body. It can protect the people in the car when the accident happens, but the blind area caused by A-pillar is enough to block the whole pedestrian, at present, there is no warning for external danger in the study of A-pillar, so a light-weight YOLOv4 blind area real-time detection method is proposed to solve the problem of external danger detection in the blind area of a-residence. Firstly, the human eye data set of CAS is collected and labeled with labelImg, secondly, a lightweight YOLOv4 detection model is proposed, and the Yolov4 backbone network CSPDARKNET53 is replaced by MobileNetv3 to reduce the parameters of the model, the deep separable convolution is introduced to replace the common convolution in the PANET network of Yolov4. Finally, the detection Head of Yolov4 is improved, and the Yolo Head is replaced by Decoupled Head. The experimental results show that the weight of the lightweight Yolov4 network is only 1/4 of that of the original network, the classification accuracy on VOC12 + 17 was 87.0% , and the detection rate was 56FPS. It can be seen that the light-weight network can be used to detect the danger of the blind area, which can provide reference for the application of the a-pillar blind area research in the actual work environment.