Research On Lightweight Convolutional Neural Network Blind Spot Detection Syste

Road traffic safety has always been an important issue on a global scale.Accidents are generally caused by driver negligence,inexperience,or visual blind spots of vehicles.For example,large vehicles have large eye positions due to their large bodies.,So that there are many blind spots around the body that cannot be observed by the driver.Pedestrian and vehicle detection technology is the most concerned part in early warning of blind spots.As a vehicle-mounted safety system,the blind spot detection system has high requirements for real-time,stability and reliability.Traditional image processing methods are not stable enough under complex road conditions and often miss detection.As deep learning has made significant progress in the field of target detection,the target detection technology based on convolutional neural network is much better than traditional target detection.However,the convolutional neural network has a complex structure and a large amount of calculation,and it is difficult to apply it to an embedded vehicle system application platform with limited resources and low power consumption.Based on the YOLOv3-tiny algorithm,this paper improves its speed and detection accuracy.The main research work is as follows:1.Proposed YOLOv3-tiny-better algorithm model.Compared with YOLOv3-tiny,it reduces the number of feature map channels and accelerates network inference;deepens the network depth,makes the extracted features more abstract and has more semantic information,and introduces a residual structure to prevent network degradation;and adds a feature fusion module Specifically,YOLOv3-tiny has only performed feature fusion once,and feature fusion is limited.We have added a first-level feature fusion module on this basis,so that the information of the deep network can be effectively fused with the shallow feature extraction layer.Classification and detection of small target pedestrians.Compared with YOLOv3-tiny algorithm,mAP has improved 12.37%compared with YOLOv3-tiny,and the model size has been reduced from 34.7MB to 2.2MB.Compared with YOLOv3-tiny,not only has the model size been reduced,but also the target detection performance has been improved.2.In order to realize the detection of the total blind area,we conducted a comparative analysis of the camera angle.In addition to considering the camera focal length,depth of field,dynamic range and other factors,according to the blind spot camera installation position,the camera’s field of view coverage modeling and imaging After analyzing the quality,suggestions for the selection of blind spot cameras are given.3.Optimize the quantization plan.The quantized model has less storage overhead and bandwidth requirements,and has a faster calculation speed.In order to save memory access and make the model faster,we chose to combine the convolution module,BN(batch normalize)module,and activation function module as our feature fusion module’s INT8 quantization scheme,and activated the scheme Function optimization and weight optimization.The experiment proves that our optimized quantization scheme has better detection performance,and the target detection speed reaches 25FPS under the embedded platform,which achieves the application of real-time detection.