Vehicle Target Detection Method Based On Deep Learning

As the most commonly used vehicle,vehicles are widely used in our daily lives.Whether in the bustling city or in the quiet suburbs,you can see a variety of vehicles on the road.However,the number of vehicles and vehicles is too large,causing many problems for the society,such as traffic accidents,traffic congestion,parking space and so on.In recent years,unmanned driving technology has developed rapidly,and computer vision has become an important link in unmanned driving technology,and there have been many breakthroughs.Using a combination of computer vision and deep learning,the researchers used convolutional neural networks to extract features from images,and used a variety of machine learning methods to classify and regression feature information,making the detection speed and accuracy of vehicle detection both.A lot has been improved.However,the current target detection method for vehicles does not well maintain the advantages of both detection speed and accuracy.For example,the Faster R-CNN detection method does not guarantee real-time effects when the accuracy is relatively good.In the YOLO and SSD detection methods,the detection speed can be ensured faster,but the accuracy cannot guarantee a good level,and more misidentifications will occur.Therefore,these kinds of target detection methods are not well applied in practice.Aiming at the shortcomings of current vehicle detection methods,this paper proposes an improved model of DMRSSD based on SSD,which improves the six feature maps for SSD detection and identification by means of multi-convolution kernel multi-receptive field,and at the same time,it is connected in a dense connection manner.The feature information of the upper and lower layers is fused together.The innovative method characteristics and research work of the improved SSDbased model DMRSSD in this paper are as follows:(1)In the six feature maps of the original SSD output,the two feature maps with larger lengths and widths use MR-Block,a convolution neural network module with multiple convolution kernels and multiple receptive fields,to extract features again,to extract features in the middle part of the feature map using smaller convolution kernels and convolution kernels sliding windows,and to extract features in the surrounding part using expanded convolution kernels and larger convolution kernels sliding windows,and then to extract different convolutions.The combination of feature information extracted by kernel makes the output feature map more rich in semantic information,that is,extracting more detailed feature information from the middle part while taking into account the feature information of the surrounding part.(2)The dense connection mode is used for the six feature maps output by DMRSSD.Like the dense connection mode of the feature maps in Dense Block,when each feature map is generated,the larger feature maps in front will be directly involved.In this way,the feature information can be reused,which makes the feature information generated later more relevant.(3)In model training,Focal Loss loss function is used to replace the original classification cross-entropy loss function,which makes the positive and negative samples more balanced and improves the accuracy of model training results.AdaBound optimization function is used to adjust the parameters of model training dynamically,which can accelerate the convergence speed of loss function and improve the accuracy of model detection.The experimental results show that the improved DMRSSD model is used to detect the vehicle target of the KITTI dataset.The mAP is increased from 82.79% of SSD to 90.57%,an i9 ncrease of 7.78%.