Research On Vehicle Detection And Recognition Based On Deep Learning

With the development of Artificial Intelligence(AI)technology,and the AI industry has greatly increased under the support of the government,many companies and institutions have invested in the research and development of AI products.For one,Intelligent Transportation System(ITS)has become one of the hotspots of research in industry and academia.Vehicle detection and recognition is an important technology for the construction of ITS.The main aspects include automatic driving,traffic-flow statistics,object tracking,and positioning,etc.It has great application prospects and research value.However,in the practical implementation,as the complexity of natural scenes,vehicle detection will encounter various challenges,such as weather,viewing angle,occlusion,and so on.In this paper,we combined with actual engineering application scenarios,we studied vehicle detection and recognition based on the deep learning method.The main research content of this article is as follows:First,briefly reviewed the history of deep learning,and the theoretical basis,implementation principles,and described the advantages and disadvantages of convolutional neural networks(CNNs).Mainly analyze the convolutional layer of CNN,several commonly used activation functions,pooling layer,and the basic network including with VGG and ResNet.Then,we have deeply studied the object detection models including the two-stage detection algorithms represented by R-CNN and the one-stage detection algorithms represented by the YOLO series and SSD.Secondly,as traditional object detection cannot quickly and accurately respond to the changing traffic environment,this paper chooses a one-stage detection algorithm to achieve vehicle detection,selecting the current best-performing YOLOv4 framework as the benchmark,combined with actual project applications,and compressing the model the goal is to increase speed and build a fast and efficient vehicle detection model.By improving the algorithm framework and data augmented strategy,the inference speed of the vehicle detection model is increased by 43% while the detection accuracy is closed to the YOLOv4.Finally,in the case,that vehicle recognition is also based on ensuring accuracy,improved the speed of the classification network.By introducing CSPNet,SE attention mechanism,Swish activation function,and other strategies to PeleeNet,the final BFLOPs were reduced by 13%,and the accuracy of TOP-5 was improved by 1%.The research results of the thesis are applied to the "Smart Parking" project and the "AI Guard" project.Among them,"Smart Parking" has completed the online test and will soon be put into production and application."AI Guard" has completed the algorithm design and the algorithm migration of the hardware platform,and the business logic is still under development.