Research On Vehicle Target Detection And Target Collision Control Of Target-hitting Robot Based On Vision

As an important branch of robotics,intelligent mobile robot integrates technologies related to perception,decision making and motion control.These technologies have good theoretical research significance and practical application value,and it's the focus of academic and industrial research.However,today's intelligent mobile robots often require higher computing power of the device.Therefore,the cost is high and the popularization of application is difficult.In view of this,aiming at the problem of efficient realization of autonomous functions by low-cost devices,this paper studies and realizes the core functions of a combat Target-Bump robot(Target detection,Target decision and Target motion control)on Jetson Nano board,based on the core technologies of mobile robots(perception,decision and motion control).In the part of Carhead and Target detection module,the deep learning target detection technology based on vision is adopted to achieve efficient vehicle target detection function for low-cost equipment.Firstly,1523 sample images are collected,and the data sets of the target body and target are established and enhanced based on the method of illumination transformation.Three lightweight detection algorithm models suitable for embedded device deployment were screened and trained as candidate models for vehicle target detection,namely Nano Det,YOLOv5 s and Yolo-Fastest V2.Then,two optimization experiments were carried out for the alternative model :(1)model pruning optimization experiment.The BN layer coefficients are constrained by L1 regularization,and the layers with very small coefficients are trimmed.(2)Model deployment optimization research experiment.Tensor RT technology is used for interlayer fusion and data quantization.Finally,the running speed of the model on the development board reaches 26.3FPS,m AP is 91.6%,and the detection effect is good.In the decision module of bumping the target part,combining the knowledge of matrix game,the rule-based method is used to design the decision-making algorithm of robot bumping target.By analyzing the process of Target-Bump,the position situation of both sides in the battle is divided into simple and complex situations.Firstly,the simple decision rules are designed,and then the complex decision payment matrix is designed based on the matrix game method,and the decision benefits are calculated by the actual experimental data.Finally,the strategy of simple and complex case is integrated,and the whole target hitting decision algorithm is established.In the movement control module,aiming at the task requirements of fast target collision,a target movement scheme of correcting deviation first and then moving forward at full speed was designed.According to the sensor principle of the camera and the kinematic model of the robot,a visual servo control system based on image is established.Aiming at the complex and changeable robot movement environment,a Fuzzy RBF neural network PID controller which can adjust parameters automatically is designed.It is compared with conventional PID and fuzzy PID controller by simulation.Finally,the control function of the robot is tested in the actual environment,and the results show that it can meet the requirements of the task.