Research And Implementation Of Vehicle Detection Algorithm Based On Depth Learning

With the rapid development of pattern recognition,computer vision and artificial intelligence,vehicle detection has drawn extensive attention as the core of pattern recognition.It is the basis for implementing an intelligent driver assistance system and even a vehicle's driverless mode.The traditional vehicle detection algorithm is to detect the moving vehicle by using the differences of symmetry,color texture and boundary features of the moving vehicle in the image sequence.However,due to the influence of the weather in the environment,such as light intensity,shade of the obstruction,the uncertainty of the target sports vehicle and the complicated background,it becomes very difficult for the traditional algorithm to detect the vehicle with high robustness and high precision.The final test results are not satisfactory.This paper starts with the traditional detection algorithm of moving vehicles and proposes some new algorithm frameworks for the difficult problems encountered in traditional algorithms.The main work and research results of the dissertation are as follows:1.In the traditional vehicle detection algorithm,the vehicle detection effect is unsatisfactory due to the mutual occlusion between vehicles,the complicated traffic scene and the low stability of the vehicle detection model.Therefore,this paper presents a vehicle detection algorithm based on Haar features and Adaboost classifier.Firstly,we use the weak classifier trained by road vehicle dataset collected by vehicle-mounted camera and then combine the best weak classifier into a strong Adaboost classifier according to a certain weight.With the help of the existing OpenCV library in image recognition,vehicle detecting goes on successfully.The test results show that the algorithm has a good detection effect on road vehicles,and break the bottleneck of the traditional detection algorithm.However,there is still a problem that the accuracy of feature labeling is not high.2.In view of the above-mentioned defects of low accuracy of manual annotation,this paper introduces the deep learning algorithms,which are pretty popular.Starting from the basic theory of deep learning,the convolutional neural network is used to extract the features of the features autonomously.In addition,it can extract the features of the images,predict the candidate regions and finally get the test results.Aiming at the shortcomings of a large number of candidate frames generated by Region Proposal Network(RPN),a new detection framework is proposed in this paper:a concatenation structure of RPN and Fast R-CNN,named Improved Region Proposal Network(IRPN)algorithm.The candidate box output by the RPN is used as the input of the Fast R-CNN to filter the candidate box once more to obtain a small number of high-quality candidate area boxes toimprove the detection effect of the vehicle.On this basis,a large number of experiments are carried out in this paper.The test results of the cascaded IRPN algorithm are tested on the datasets collected by the author,and good detection results are obtained.The feasibility and effectiveness of the improved cascade framework are proved.3.In order to identify and locate the vehicle more quickly and accurately,an algorithm based on convolutional neural networks is proposed in this paper.The algorithm is optimized in the proposed network and feature extraction.Specifically,our framework is embedded with a lightweight proposal network(LPN)to generate initial anchor boxes as well as to early discard unlike regions.Feature fusion technology is used to extract Hyper feature and refine the identification and location of the vehicle,which improve the quality and accuracy of vehicle detection effectively.We evaluate our network on the recent Large-scale data sets and the vehicle data sets collected by ourselves.We obtain a significant improvement over the state-of-the-art algorithms,especially the Faster R-CNN by 9.91% mAP(mean average precision),which fully shows the effectiveness of the algorithm.