Pedestrian-vehicle Detection Model Based On Optimized YOLOv3

The pedestrian-vehicle detection task is an important part of intelligent transportation and intelligent road.Whether it is automatic driving or road condition monitoring,it will bring great convenience to people’s travel in the future.YOLOv3 is one of the most effective algorithms in the current target detection task.Based on the original YOLOv3 algorithm,this paper optimized and improved its shortcomings to improve results in the pedestrian-vehicle detection task,and verified the results and compared the indicators on the data set.The main work of this thesis is as following:1.This thesis elaborates on the theoretical basis of convolutional neural networks,including its basic structure,commonly used operations in the network,etc.,and it explains why convolution neural network achieved a unique advantage in the field of target detection;The structure,principle and prediction process of the original YOLOv3 algorithm are introduced in detail.2.Optimize the original YOLOv3 algorithm in terms of network structure.According to the shortcomings of the original algorithm model,add two stages for detection,and adopt the method of cascading different scale feature graphs to improve the transmission efficiency and utilization of the original network features;In terms of deep network receptive fields,the introduction of a dilated convolution structure ensures that the feature map has a larger receptive field while ensuring its resolution and improving the degree of information restoration;in the up-sampling stage,the feature pyramid structure of the original model is improved To reduce the burden on the network to achieve better detection results.3.In the loss function and the non-maximum suppression on the original YOLOv3 algorithm optimization,through the introduction of the calculation analysis of GIo U’s advantage,and use it as the regression loss function,improve the training of bezel convergence speed and the accuracy of the prediction.It also will be used as the non-maximum suppression index,increase the ability of the model for frame filter slightly,further improve the final detection accuracy.4.The optimization scheme of this article is verified by experiments,and the KITTI data set used and the evaluation indicators used in the task of pedestrian-vehicle detection are introduced.Test the trained network under different improvement schemes,and compare the indicators and display the detection effect.The final result shows that the algorithm optimized in this paper has a certain degree of improvement in detection speed and accuracy compared with the original algorithm.