Research On Detection Methods Of Vehicles And Pedestrians Based On Convolutional Neural Network

Traffic object detection is not only the key but also the difficulty to realize unmanned driving,and its accuracy and real-time performance are important indicators to evaluate the detection system.At present,application systems based on deep learning technology are mainly deployed in cloud servers,the network environment will greatly affect the response speed of the system.Deploying the system in a small embedded computer is more conducive to the improvement of the system speed,providing stable services and protecting the privacy of users.However,with the improvement of the object detection accuracy,the parameter storage cost and calculation amount of the model are also increasing.It is of great practical significance to choose a object detection algorithm with high precision,low memory requirement and less computation.Therefore,on the premise of ensuring the accuracy,thesis proposes a solution to reduce the size of the model and memory requirements of traffic object detection.The work of this topic is summarized as follows:Firstly,thesis describes the research background and significance of this topic,and analyzes the research status,development trend and existing problems of object detection at home and abroad.Thesis studies the methods of object detection proposed by some scholars,and introduces and analyzes the current mainstream algorithms: DPM algorithm,Faster R-CNN algorithm,SSD object detection algorithm,etc.Select and make effective data sets for the experiment,and select effective evaluation indicators to quantify the experimental results.Then,based on the analysis of current mainstream object detection methods,two network compression methods are proposed for the object detection algorithm of YOLOv3 with better performance at present:(1)An improved model YOLOv3-M based on lightweight network is proposed.In this model,the depth separable convolution is used instead of the traditional convolution to reduce the amount of parameters and the calculation of the convolution process,and the inverted residual structure is used to increase the feature dimension to improve the convolution effect.In order to improve the accuracy and speed up the convergence of the algorithm,the K-means++ algorithm is optimized to re cluster the object frame,and the intersection and union ratio is used as the distance evaluation standard.SS-NMS(Soft-NMS+Softer-NMS)method is used to select and optimize the box proposal,and the final detection results are obtained.Experimental results show that compared with the mainstream object detection algorithm,the detection speed of YOLOv3-M is greatly improved,reaching 78 frame/s,which is 3.5 times of YOLOv3;in terms of the mAP index,compared with the YOLOv3 model with the best detection effect,the model is only reduced by 3.32%.Moreover,by using the new non maximum suppression algorithm,the mAP of YOLOv3-M is improved by 5.84%,and the prediction results are optimized.(2)In order to improve the real-time performance of the object detection algorithm and reduce the memory requirement,the model of YOLOv3 based on pruning is proposed.First of all,by analyzing the structure of YOLOv3 network,the pruning convolution layer can be divided into two types: ordinary convolution layers and convolution layers of cross layer connection.For the pruning of ordinary convolution layer,channels with less information and redundant information are pruned by using L1-norm and FPGM double evaluation index;for convolution layers of cross layer connection,the group pruning method is used to solve the constraint pruning problem and the scaling factor is used as the evaluation index.After pruning,the network m AP decreased by only 2.26%,the memory demand decreased by 54.40%,and the FLOPs decreased by 55.06%.Finally,in order to verify the practicability and reliability of the algorithm,the pedestrian and vehicle object detection system is designed and implemented based on the first method,which includes two modules: image object processing and video object processing.The experimental results show that the system can achieve object detection well.