Research And Implementation Of Vehicle Detection Algorithm Based On Deep Learning

With the development of computer science,communication technology,and automatic control theory,automatic vehicles have step from science fiction to reality.Many technology companies,including Google,have shown unprecedented enthusiasm for the autopilot technology involved in the field of the automatic vehicles.Therefore,the research on vehicle detection algorithms has become a hot research field in the realization process of automatic vehicles.In recent years,with the development of GPU technology,the theoretical research of deep learning algorithms has risen to a higher level,and some image detection algorithm innovations based on CNN deep neural network structure have emerged.Compared to traditional machine learning algorithms,those innovations bring huge detection accuracy improvement.However,due to the large scale of network,those deep-learning based algorithms are difficult to implement in the embedded system.They have the disadvantage of computationally and financially high cost which cannot meet the application requirements of real-time vehicles detection.This thesis firstly aims to improve the performance of vehicle detection algorithms based on deep learning.On the one hand,to meet the requirements of scene picture for vehicle detection tasks and optimize data set,this thesis proposes to reconstruct methods of the data set used by the deep learning algorithm based on error curve analysis method and error table analysis method.On the other hand,the key parameters in the deep learning algorithm are compared and selected,and a set of algorithm parameters most suitable for the vehicle detection task are extracted.Through the above two aspects of research and experiments,the optimization of deep learning algorithm has been done and improves of detection accuracy.Secondly,focusing on the problem of slow running system speed and high hardware cost due to excessive calculation of embedded deep learning algorithm in vehicle detection embedded application,an optimization scheme for structural cutting of deep neural network is proposed.By studying the relationship between network performance degradation and network cropping ratio,combined with the problem complexity analysis of vehicle detection tasks,a miniaturization of deep learning algorithms is realized.An X-TINY YOLO network structure suitable for vehicledetection embedded implementation is proposed,which greatly reduces the computational complexity at the expense of a small amount of detection accuracy decline.This network structure makes it possible to implement real-time embedded vehicle detection based on deep learning algorithms.Finally,considering the parallel computing structure existing in the deep neural network,the micro ARM+FPGA hardware architecture is selected to design and implement a real-time vehicle detection system with a cost within only one thousand dollars.In the implementation,the problem of data split transmission and pipeline calculation on the limit of hardware resources is mainly solved.The system is tested by vehicle drive test and has been proved meeting the real-time requirements of the vehicle detection task along with high detection accuracy.In summary,this thesis based on the research of deep learning algorithm optimization and its application in the field of vehicle detection,finally realized a low-cost real-time vehicle detection system which meets the requirements of real-time and detection accuracy.This study provides a complete scheme to the achieve vehicle detection of autopilot technology involved in the field of the automatic vehicles.