Driver Seatbelt Detection Based On Deep Learning

With the increasing number of motor vehicles,the problem of road traffic safety is becoming more and more serious.When a traffic accident occurs,the driver wearing seat belt can effectively protect the driver's personal safety.Relying on manual detection of traffic surveillance images is the main way for the current public security management departments to supervise drivers wearing seat belts.With the development of artificial intelligence,the intelligent system based on automatic image recognition has been deeply studied and applied in different fields.Aiming at the technical requirements of automatic identification and detection of whether drivers wear seat belts or not,this paper studies small target detection and recognition methods and their applications in traffic surveillance images based on deep learning.In this paper,the research status of deep learning and convolution neural network and its theory and technology are deeply analyzed and studied.Deconv-SSD,Squeeze-YOLO and semantic segmentation based driver's seat belt detection algorithm are proposed.Deconv-SSD achieves vehicle detection quickly through deepthwise separable convolution and fusion of multi-resolution feature maps.Then the driver area can be located quickly by Squeeze-YOLO algorithm using the features of front windshield and lightweight feature extraction.In the location area,the driver's seat belt is quickly segmented based on semantic segmentation algorithm and pruning technology,and the driver's seat belt is detected by judging the maximum connected area of the segmented seat belt.In this paper,experiments and data analysis of the proposed algorithm are carried out.When the image resolution is consistent with the feature extraction model,the mAp of Deconv-SSD is increased from 77.2% to 79.6% in PASCAL VOC data set compared with SSD algorithm.In the self-made seatbelt detection data set,Squeeze-YOLO achieves 73 FPS at 99.96% mAp,and the semantic segmentation algorithm achieves 94.87% accuracy at 305 FPS speed after pruning acceleration.Based on Cafe framework and Qt graphical interface library,this paper completes the design and implementation of driver's seat belt detection system,and achieves the engineering design of GPU acceleration algorithm by combining batch normation layer and Tensorrt quantitative model.Finally,this paper designs a lowpower and low cost seatbelt detection algorithm based on XILINX open source IP core,and implements it on embedded Soc based on FPGA.