Vision-based Robot Target Detection And Trajectory Prediction Research

With the continuous development of the modern intelligent manufacturing industry,industrial robotics is widely used in spraying,welding,assembly,aviation,and other industrial manufacturing industries,and for complex operational tasks,robots need to collaborate with operators to complete the task.In the process of manmachine cooperation,the safety problem caused by collision becomes the focus of its development.In order to prevent damage to humans or robots caused by collision,the robot needs to have a collision detection function.In this topic,a vision-based collision detection method is designed to address the safety issues in the process of human-robot collaboration,to detect the target of the operation process,and to predict the trajectory of pedestrians to determine the possibility of collision between pedestrians and robots,so that the robot can respond safely.The main research contents are as follows:For the subject target detection and trajectory prediction research requires the use of Kinect 2.0 depth binocular camera as a visual sensor to obtain information,based on the mathematical imaging of depth binocular camera and the principle of camera aberration to establish its mathematical model.According to the principle of camera color lens and depth imaging lens marking,Zhang Zhengyou’s calibration method is used to conduct calibration tests for tessellation grid pictures taken by the depth camera from different angles and solve to get the specific parameters of the camera.The classical target detection algorithms such as the frame difference method,background subtraction method,and optical flow method are analyzed,and the background inter-frame difference method is used for target detection of video sequence information,and the selection of threshold value and update of background model is optimized for the problems of false detection,background occlusion and background color similarity in the target detection process,so that it can get the motion target model with high accuracy,and the target detection process exists the image processing methods are analyzed,and the captured images are experimented with the background subtraction method and the optimized background inter-frame difference method,and the comparison reveals that the phenomenon of voids in the model is reduced,and the optimization of the threshold value reduces the false detection phenomenon,and a more complete target outline can be obtained.The image information in the target detection process is preprocessed,the location coordinate information of the pedestrian is extracted by the depth image,the human skeleton information is obtained by using the Kinect 2.0 depth camera,the bounding box of the pedestrian with a rectangular is determined,the head orientation information and feature information are obtained by feature extraction of the human skeleton information,the pedestrian intention prediction model is established,and the combination of Long Short-Term Memory neural network algorithm to predict the trajectory of the pedestrian,analyze the future trajectory of the pedestrian,determine whether the intersection of the bounding box and the robot’s workspace produces a hazard assessment to determine whether a collision occurs,and make a safety response decision.On the basis of considering the generality and practical economy of the algorithm,a two-degree-of-freedom robotic arm test platform was built,and experiments on target detection and trajectory prediction during the operation of the robotic arm were carried out.During the experiment,the system was able to effectively detect and track the target of the working environment in real time,obtain a high-definition motion target model,and make trajectory prediction for the target.