Object Detection Of Intelligent Driving Based On Lightweight Neural Network

Traditional target detection algorithm applied to intelligent driving,always adopts two-stage algorithms.With subsequent tracking and reasoning strategies,the algorithms can guarantee high accuracy.This type of algorithm has the advantage of stable performance,but the calculation speed is slow,and it has high storage requirements,so the requirements for hardware and memory space are very high,and it is difficult to achieve mass production.For intelligent vehicle object detection,it is necessary to consider not only the accuracy,but also the inference time and the number of parameters when designing the network.So,how to improve detection speed as much as possible on the premise of ensuring accuracy is a topic worthy of studying.Aiming at the above problems,this paper carries out the research of intelligent vehicle object detection based on lightweight neural network.The main research contents are as follows:（1）Overview of object detection algorithmsFirstly,the principles of common object detection algorithms are analyzed.Secondly,the evaluation metric of object detection algorithm is discussed.Then,the mathematical principle of neuron model,which is the basis of neural network,and some basic structures of convolutional neural network are analyzed in detail.Finally,four kinds of loss functions,neural network training and regularization are introduced.These include L1 and L2 regularization,Dropout and Batch Normalization.（2）Object detection in road scenes based on improved YOLOv3The YOLOv3 object detection algorithm has been improved to increase the accuracy of object detection and make it suitable for road scenes.First,YOLOv3 needs to manually set the sizes of the anchors.This article redefines the distance formula,replacing the original Io U（Intersection over Union）with GIo U（Generalized Intersection over Union）and clusters the BDD100 K data set to get the anchor sizes suitable for the road scenes;Secondly,add a new data augmentation method——Cut Mix,which cuts and pastes between training images,where the labels are also mixed proportionally.At the same time,we improve the feature fusion part to fuse features at a lower level to increase the detection accuracy of small targets;Finally,the results of the comparative experiments are analyzed in detail.It shows that the detection accuracy of YOLOv3 algorithm is increased by 6.2 percentage points by using the above methods.For small goals,the improvement is more obvious,with an average increase of 13 percentage points.This model will become a teacher network in knowledge distillation.（3）Road scene object detection based on lightweight neural networkLightweight the improved YOLOv3 network with higher detection accuracy to improve inference speed and reduce memory usage.First,draw on the idea of YOLOF（You Only Look One-level Feature）,and perform structural lightweight processing on the improved YOLOv3 network that has been obtained.YOLOv3 performs classification and regression on three feature maps at the same time,and this article will use dilated convolutions to gather all semantic information in one feature.This way can achieve or even exceed the detection effect of YOLOv3;Secondly,the training method of knowledge distillation will be adopted.Use the improved YOLOv3 network as a teacher network and the lightweight network as a student network,which can further improve the detection accuracy of the lightweight model;Finally,use the method of ablation experiment to compare and analyze the above methods.The test results show that the m AP（mean Average Precision）value of the model obtained through the above methods will be 12 percentage points higher than that of the original YOLOv3,and the fps（frames per second）is 3 times that of the original YOLOv3.At the same time,this article compares our model with one-stage and two-stage classic models and the latest object detection algorithms,the accuracy and speed of our model are better than classical models.For the latest detection algorithms,the precision of the model in this paper is slightly lower but gets higher detection rate,so we can say the model reach the balance of the speed and accuracy.