A Method Of Robot Tracking Based On Deep Learning And Monocular Vision

Human beings can judge the category of external environment things by vision and roughly perceive the distance between themselves and objects,so as to achieve the purpose of following or avoiding.For robots,the traditional AGV navigation methods,such as magnetic wire,magnetic nail,two-dimensional code and laser,can not meet some of the needs of people in life and production because of their high hardware cost and low intelligence.Visual system has been widely used in autonomous positioning and navigation of robots in recent years.Monocular vision has strong practicability in short-range navigation because of its high maturity,low cost and strong real-time performance.This paper proposes a method based on deep learning combined with monocular visual ranging to achieve robot tracking.The internal and external parameters of the monocular camera are obtained through the calibration process.The SSD algorithm in deep learning is used to identify the target and obtain the pixel height of the target frame.Then calculate the distance between the camera and the target object according to the similar geometric principle in the monocular ranging algorithm.Finally,the robot's autonomous tracking is completed by sending a Twist message to the robot base.This is the main idea throughout this article and main tasks are as follows:(1)In the process of monocular ranging,in order to use the principle of geometric similarity to obtain the similar distance,it is necessary to obtain the internal and external parameters of the camera itself.Accurate acquisition of the focal length of the camera has also become an important step in the experiment.Therefore,this paper uses Zhang's calibration method to calibrate the camera,improve the calibration accuracy,and prepare for the follow-up work.(2)In the process of robot tracking,the first step is to identify the target to be tracked.In this paper,SSD target detection algorithm based on deep learning is used for target recognition.Compared with the traditional target detection algorithm,the algorithm greatly improves the accuracy and speed.This algorithm has high robustness to scale change.It uses the convolution features of multi-layers,different sizes and sensing fields to recognize targets.(3)After detecting and recognizing the target by SSD algorithm,the pixel height of the target in the image is calculated.Using linear monocular vision ranging model and geometric similarity,the relationship between the object and the imaging object can be determined in proportion.The model selected in this paper is more convenient and simple to realize ranging.The distance can be measured quickly if the necessary parameters are obtained.Although it can only be used to measure similar distance,it can not get the coordinates of each point measured,but it is more in line with the particularity of tracking robot.(4)After obtaining the distance between the robot and the target object,in order to realize the autonomous movement of the robot,this paper adds a critical value to the measured distance value.When the measured distance is larger than this value in real time,the control command Twist is issued to the subject of cmd_level through move_base,giving the left and right wheels a positive speed in the direction of the target.Similarly,when the measured distance is less than the critical value,a negative velocity is given to the left and right wheels.This way of issuing Twist commands can make the control process of the robot very fast and simple.