Application Of Improved Faster-Rcnn In Vehicle Detection

Vehicle detection is a hot research topic in the field of computer vision,and it is also an indispensable part of perfecting intelligent transportation systems.Traditional machine learning methods use artificially designed target features.In complex application environments,the algorithm has low robustness and weak generalization ability.The convolutional neural network can learn the target features autonomously,and it can still ensure high accuracy in the face of complex and changeable environments.Therefore,the application of the convolutional neural network method for vehicle detection has been extensively studied and is also of great significance.The research contents of this science article are:(1)Analyze the target detection algorithm using traditional machine learning methods and convolutional neural network methods,and then focus on the detection principle of Faster-Rcnn network,which has an important position in target detection;(2)By analyzing the parameters of commonly used convolutional neural networks,it is proposed to replace the VGG16 network with a residual structure or an Inception structure with fewer parameters.(3)The depth of the convolutional neural network continues to deepen,which reduces the size of the feature image,which eventually leads to serious loss of target features,and then the target is missed.The shallow semantic information contains rich target location information.A method of feature fusion and multi-scale prediction is proposed to solve this problem.(4)The convolutional neural network has a long training time and slow convergence.In order to shorten the training time,the samples in the vehicle data set are clustered using the K-Means algorithm to regenerate the anchor size suitable for the vehicle.By experimenting with the improved Faster-Rcnn network on the KITTI data set,the results show that:(1)A network structure with fewer parameters reduces the detection time of the model,while improving the detection accuracy of simple samples in the data set;(2)After fusion and multi-scale prediction methods,the accuracy of vehicle detection has been greatly improved compared to the original network.Especially after using the method of feature fusion and multi-scale prediction,the accuracy rates of simple samples,medium complex samples and complex samples in the KITTI data set have been improved by 6.25%,9.3%and 10.91%,respectively.This result shows that for occlusion or small targets,feature fusion and multi-scale prediction methods can effectively improve accuracy.(3)Although resetting the anchor point size does not improve the accuracy of the model,it makes the model converge faster during training.