| With the development of computer vision,target tracking and other technologies,handheld stabilizers with ordinary mechanical structures are increasingly unable to meet people’s needs for better assistance in shooting.Establishing an intelligent visual system is an important step in the intelligence of handheld stabilizers,whose main function is to automatically detect and track the images captured by the camera,helping the photographer to better complete the shooting work.However,generally speaking,the captured scenes are in a mobile state and are affected by factors such as distance and network environment,making it impossible to use computers or networks to detect and track targets.However,building an intelligent vision system directly on an embedded platform can effectively solve this problem.Therefore,this article starts with embedded technology,combines object detection and tracking technology,and designs and implements an embedded object detection and tracking system for handheld stabilizers to meet the automatic detection and tracking requirements of the stabilizers for shooting targets in certain environments.The main research content includes the following aspects:(1)Existing object detection models based on convolutional neural networks often require high computational and storage capabilities.Therefore,a structured pruning method based on convolutional neural networks is proposed to compress the model for deployment on embedded devices.Adding an intermediate graph feature analysis framework after the convolutional layer,calculating the importance of the convolutional kernel,and pruning the convolutional kernel under the threshold can achieve compression of the model.(2)A YOLOv5 s Embedded object detection framework that is more suitable for embedded device deployment is proposed in the object detection section of the stabilizer intelligent vision system.Firstly,using the idea of Mobile Net deep separable convolution,modify the YOLOv5 s backbone network structure.At the same time,combined with structured pruning methods,further compression is carried out to obtain an object detection network that is more suitable for running on embedded devices.(3)In the target tracking part of the stabilizer intelligent vision system,an improved KCF target tracking algorithm is proposed based on the KCF target tracking algorithm and combined with the target detection network.Firstly,the target detection architecture is used to detect the video stream and obtain the features of the target that need to be tracked.Then,the KCF algorithm is used for tracking.When the tracking target encounters issues such as occlusion or loss,a detection algorithm is used to re detect the image,determine the tracking target through the maximum Io U(Intersection over Union),and refresh the tracking area to continue tracking.(4)In order to verify the effectiveness of the proposed target detection and tracking algorithms,a visual system verification platform for handheld stabilizers was built based on the practical application scenarios of handheld stabilizers.In terms of hardware,various parts such as the servo system have been selected.In terms of software,the overall system architecture was designed,and the web system development was achieved using Flask and VUE frameworks.At the same time,the real-time object detection and tracking interface was displayed in combination with hardware.Finally,experiments and analysis were conducted on the system in this paper,proving its practicality. |
