Research On Visual Tracking Technology For Mobile Robot In Natural Scenes

With the improvement of computer hardware level and the development of deep learning,computer vision technology has been greatly developed.Vision-based target tracking is an important research direction in the field of computer vision.There are great demands in many fields such as service industry,industry and special industry.One of the application scenarios is that in the natural scene,when the mobile robot tracks the task at the beginning,the tracking target is randomly designated to perform non-contact long-term tracking on the target.For the analysis of this scenario,the difficulty lies in the need to achieve long-term stable tracking for different targets and in the case of changes in the external environment.The existing target tracking algorithms are mostly suitable for short-time tracking.It is difficult to achieve the ideal tracking effect in the case of target size change and background interference.Based on this,this paper proposes a method of combining target tracking and target detection to accomplish target tracking.The anti-interference ability of the method is enhanced,and stable tracking can be achieved even when the target is deformed,a small part of the occlusion,the light and shadow changes,etc.,and no cumulative error is generated,which is more suitable for long-term target tracking.The main work and innovative research results of this paper are as follows:(1)Improve the SiameseFC target tracking algorithm,predict the target position,better target the fast moving target,and design a more flexible model update strategy.When the target may occlude or change greatly,reduce the template update.(2)Introducing the target detection algorithm,for the scene of mobile machine vision tracking,based on the yolo v3 framework,the target detection model is retrained,and similar categories are merged to reduce the number of classifications.(3)Combining the target tracking result with the target detection result,a target location correction relocation mechanism is proposed.This mechanism combines the high real-time performance of the target tracking algorithm with the context information utilization ability and the classification and positioning ability of the target detection algorithm,and obtains a more robust tracking method.(4)Establish a camera model and calibrate the camera.The distance and orientation of the target and the robot are obtained by the camera model and the tracking algorithm.The autonomous following task for the target is completed on the mobile robot platform using Kalman filtered distance and bearing data.