Real-time Target Tracking And Pose Measurement Of Robots In Complex Environment

It is an important prerequisite for the autonomous operation of intelligent robots obtain the real-time pose information of the object.It’s a problem badly needing to be solved that how to deal with interference from complex scenes,track targets stably and accurately in real time,and then measure required information.In this thesis,a framework to implement real-time target tracking and pose measurement for robots in complex scenes is proposed.The three aspects of light treatment,object tracking and pose measurement are studied respectively.The specific work is as follows:(1)Aiming at the influence of complex light on images,a method of light treatment based on the atmospheric scattering model is proposed.Based on the atmospheric scattering model in meteorology,the mechanism of the influence of light on the image is analyzed and then the model of recovering the original clear image from the affected image is deduced.CNN is used to fit the parameters related to the input in the model to realize parameter estimation that can adapt to the input.The above-mentioned end-to-end light treating model is embedded into the entire algorithm framework for image preprocessing,providing a high-quality input for subsequent tasks.(2)Aiming at the problem that the traditional visual tracking algorithm is difficult to deal with the scale and posture change and occlusion of the object in the image,and the need to manually input the object position in the first frame,a fully automatic object detection and tracking method is proposed.Perform target detection on the first frame based on Faster RCNN and use Dense Net instead of the default Res Net in the network as the feature extraction module which takes advantage of the dense block to maintain information transmission when the network is too deep.The accuracy of object detection in the first frame is improved as well as.The detection results are input to the object tracking module together with subsequent frames.Based on the Siam RPN network,the feature pyramid network is introduced,and the feature maps output by the feature extraction module in Siam RPN are cross-level concatenated to obtain multi-scale features.Under the premise of ensuring the speed of the model,the accuracy of tracking and positioning is improved,and it can adapt to the dynamic changes of the object scale and posture.The result of object tracking will be used as the processing area for subsequent pose measurement,so it avoids most of the environmental interference information.(3)Aiming at the problems of instability and insufficient accuracy of traditional pose measurement algorithms,a pose measurement method based with two-dimensional markers is proposed.Design two-dimensional markers with specific patterns and fix them to the appropriate position of the object,and convert the pose measurement problem of the target into the pose measurement of the two-dimensional markers.Based on the quadrilateral fitting and the pattern information of the markers,the markers are searched and screened out from the image,and their regions are segmented.The pose information is calculated based on the camera model and Pn P algorithm,and the pose information of the target relative to the end of the robot is obtained through the transformation of the hand-eye relationship.Taking the working environment of a live working robot as an example,selecting lightning arresters and wire clamps as the objects to verify the method proposed in this article.Experiments show that the algorithm framework proposed in this thesis can track targets stably in a complex environment,and measure pose information accurately in real time,which has the value of popularization and application.