Front Vehicle And Pedestrian Detection Algorithm Based On Lightweight Neural Network

The detection of road objects is the premise for completing the task of autonomous driving in traffic environment.With the booming of vision object detection,the accuracy and speed of detection are quickly improved.But for intelligent vehicle object detection,it is necessary to consider not only high accuracy,but also low delay.So,it is an important study to balance the accuracy and speed of intelligent vehicle for object detection.Aiming at the above problems,this paper carries out the research of front vehicle and pedestrian detection based on the lightweight neural networks.Then the main research contents are as follows:(1)The object detection algorithm is analyzed theoretically,and the construction of data set are introduced.Firstly,this thesis analyzed the main model of object detection and lightweight convolution module.At the same time,this paper collected a series of data to constructing actual traffic road environment data set,and expanded the number of labels for the category with little samples by data enhancement.Then this paper used improved K-means algorithm to cluster the anchor for data set.(2)A lightweight object detection based on improved YOLOv5 is built.According to the performance in collected data set for mainly object detection framework,YOLOv5 is selected as the benchmark model of this paper.For the slow detection speed and redundant parameters,this paper reconstructed the model with light neural network,used the advantage of highly resolution network and fused the spatial attention and channel attention algorithm.So,a lightweight network named HR-YOLO is built.In our data set,the precision and inference time on our model are excellent,comparing with other lightweight detection models.(3)A Lightweight object detection based on model channel pruning is built.Model compression has been conducted for the benchmark model.The scaling coefficient of BN layer is used to evaluate the importance of the channel,and a pruning strategy for base model named crossing stage local network is designed to prune the redundance channel.Then the precision of pruning model was restored by knowledge distillation and fine-tuning method.On collected data set,this compression algorithm ensured excellent accuracy and speed by greatly simplifying the model.(4)An application for lightweight network of front vehicle and pedestrian detection has conducted.The lightweight model has deployed on the Wed-side and mobile-side in the paper.First of all,the man-machine interactive detection page is designed to complete the Web-side deployment of the model.Secondly,in order to verify the lightweight effect,the mobile device is equipped our model.Finally,the experiment shows that the lightweight model runs smoothly and performs well on the microprocessor.