Research And Implementation Of Target Detection And Tracking Based On Embedded Hardware And Software Co-design

Target detection and tracking is the basic content and research hotspot in the field of computer vision.With the rapid development of artificial intelligence,there is an upsurge in the research of target detection and tracking technology in different fields.Due to the interference of camera jitter,similar target confusion,illumination change,motion blur and other complex factors in the actual application scene,it is very difficult for the detection and tracking task.Target detection and tracking is usually implemented on PC.With the improvement of edge computing platform performance in recent years,especially the emergence of heterogeneous embedded systems based on ARM and FPGA,it has promoted the in-depth research and extensive application of the target detection and tracking technology based on the on-chip system,because of its advantages in power consumption,volume,cost and real-time.Due to the limitation of computing resources in embedded system,a large number of detection and tracking algorithms with high accuracy are difficult to be implemented in embedded system or meet the real-time requirements.In this paper,a target detection and tracking algorithm based on attribute recognition is proposed,furthermore,it was implemented in embedded system.The main work of this paper is as follows:(1)The current computer vision field needs to extract more semantic information from the image.In addition,in order to avoid the confusion of multiple similar targets in the detection results,this paper introduces the attribute recognition algorithm in the detection and tracking algorithm to screen out specific targets.Due to the limited computing resources of the embedded system,a lightweight multi-task which shared feature extraction network algorithm is designed.In order to improve the overall recognition accuracy of the algorithm,the design of loss function is optimized.In the loss function,the loss weight of different tasks is determined according to the classification difficulty.Experimental results show that the model of attribute recognition algorithm has smaller volume and calculation,which can meet the real-time requirements of embedded system,and the algorithm has higher recognition accuracy,and can effectively distinguish similar targets.(2)In the kernel correlation filtering algorithm,due to the lack of effective detection mechanism,the model is easy to be polluted in complex environment,resulting in error accumulation and tracking failure.This paper proposes a detection mechanism of tracking effectiveness based on perceptual hash algorithm,which can judge whether the current tracking result is effective in time and avoid tracking failure.(3)In the target detection model,only the current frame is detected independently,therefore,it is easy to miss the detection in the harsh environment.In this paper,we propose a strategy to prevent the missed detection.On the basis of kernel correlation filtering algorithm,the Lab color feature after image enhancement is introduced to effectively reduce the rate of missed detection..(4)The algorithm in this paper needs to be deployed to the embedded system finally.In order to improve the real-time performance of the algorithm,the software and hardware codesign is adopted in combination with the features of Zynq chip.Because of the large calculation of convolutional neural network and the high requirement of memory read bandwidth,this paper firstly reduces the volume and calculation amount of the target detection model and attribute recognition model through quantitative compression,and then deploys the parts with high degree of parallelism to FPGA for implementation.The other parts with small calculation amount in the algorithm are implemented by ARM.The advantages of ARM and FPGA are given full play.Experimental results show that the proposed detection and tracking algorithm based o n the fusion of attribute recognition can achieve stable tracking of the target in complex scenes,with the accuracy and success rates of 91.19% and 84.79%.Compared with KCF tracking algorithm,it improved by 21.48% and 25.85 respectively.In the quantization process of SSD model and attribute recognition model,the accuracy loss is less than 1%,which has no effect on the whole algorithm.The frame rate of the embedded hardware and software system is 23.5 FPS,which basically meets the real-time requirement of the algorithm.