Research And Application Of Object Tracking Algorithms Based On Embedded System

With the development of computer hardware,we can use large-scale convolutional neural netw ork to accomplish various tasks in the field of computer vision,such as image classification,object detection,etc.,and the performance is much better than traditional algorithms.However,in the field of deep learning,people have formed a dependence on high-performance hardware.Without these powerful computing pla.tforms,it seems difficult to get excellent models.And with the development of smart phones,wearable devices and other miniaturized devices,people hope to enjoy the convenience of artificial intelligence anytime and anywhere with the help of these embedded platforms.However.the current excellent neural network model is not suitable for small embedded devices with limited computing and power consumption.In view of the above problems,this paper starts from the classical algorithms in the field of computer vision,combines with the convolution netw ork algorithm and various model compression methods which have developed rapidly in recent years.and explores the feasibility of achieving excellent object detection and tracking algorithms on embedded devices.The main work of this paper is as follows:Firstly,this paper studies the commonly used object detection and tracking algorithms,also compares and analyzes the difference of various object detection algorithms.At the same time,several popular model compression methods is introduced in this paper.Then,this paper decomposes the YOLOv3 model,which effectively reduced the model's parameters.Then we used the pruning method to compress the YOLOv3 model one more time,which removed the redundant parameters in the model.At the same time.a 1×1 channel pruning method based on L0 norm and L1 norm is proposed in this paper,which achieves good compression performance while ensuring the accuracy of the model.Thirdly,we analyzed each part of Deep-SORT target tracking algorithm,and proposed an improved Deep-SORT algorithm based on LDA algorithms,which aims to extract appearance features.After that,we used YOLOv3 under the secondary compression and improved Deep-SORT algorithm,which runs on the embedded platform Jetson TX2 and achieves real-time performance.Fourthly,this paper applies the embedded object tracking algorithm on intelligent monitoring,so that the system could count road traffic.And all the computing tasks of the algorithm are deployed on the edge embedded platform,and we used GSM network to communicate with the embedded platform,which can run independently from the data center and has practical value.