Research On Improved Vehicle Detection Method Based On YOLOv3 Network Model

Intelligent Transport Systems is a hot research field in recent years,and vehicle detection is an important part of it.At present,the problem of vehicle detection is transformed into target detection problem in image,and target detection method in deep learning has been applied to vehicle detection,but such methods still face the problems of achieving small target detection and model lightweighting.The YOLOv3 network model is a typical deep learning target detection method,which implements feature extraction,candidate box classification and regression in the same branchless deep convolutional network to obtain better detection accuracy and speed balance,but there are still the following drawbacks: 1.The feature map used for detection does not retain the small target objects and clear edge information of objects,causing the problems of missed detection of small targets and inaccurate positioning;2.The structure is bloated,the backbone network part adopts 53-layer deep convolutional neural network,and with a residual structure added,which causes huge resource consumption and high requirements on the hardware platform,limiting the practical application of the model.Aiming at the above two issues,this article took vehicle detection as an example to carry out research.The main research work was as follows:First,in response to the problem of inaccurate positioning and missed detection of small targets when the YOLOv3 network model detects vehicles,a fine-grained detection module was added to the network.First,the deep feature map was amplified by upsampling in the original network detection part to obtain the semantic information of small target vehicles;then fused the feature map obtained by upsampling with the feature map of the same size of the shallow network;finally,the fused feature map was convolved 7 times to complete vehicle detection.Second,in view of the computational overhead and memory consumption problem of the YOLOv3 network model,depthwise separable convolution was introduced in the the network residual unit to achieve a lightweight model.The YOLOv3 network model,which was a deep convolutional neural network with many layers and parameters,large computational overhead and large network size.The depthwise separable convolution splited the original standard convolution process into depthwise convolution and pointwise convolution to achieve spatial and channel separation of the convolution process.Third,designed the SDP-YOLOv3 network model vehicle detection method.Based on the YOLOv3 network model,this method doubles the second residual module of the network,and realized effective feature extraction by deepening the network depth;fused deep semantic information with shallow representation information to obtain feature maps that retained small target information,and detected the feature map;the lightweight model was realized by removing redundant modules and introducing depthwise separable convolution;finally,the hyperparameters were fine-tuned to make it suitable for the research scenarios in this paper.The experiments showed that the designed SDP-YOLOv3 network model was able to detect small target vehicles with a size of 2% of the original image size,and the AP(Average Precision,AP)was improved from 93.93% of the YOLOv3 network model to 97.57%;in terms of model scale,the SDP-YOLOv3 network model was reduced to 79.3 MB(MByte,MB),which was equivalent to 31% of the YOLOv3 network model;in terms of computing complexity,the SDP-YOLOv3 network model was reduced to 34.814 BFLOPs(Billion Foat Operations,BFLOPs),which was equivalent to 53% of the YOLOv3 network model.Generally speaking,the SDP-YOLOv3 network model improved the AP based on the YOLOv3 network model and reduced the model scale and computational complexity.The research work in this paper would be informative for convolutional neural networks to achieve small target detection and achieve lightweight model.