Research On Object Detection And Tracking Technology Based On The MDP Model

Vision is an important channel for human and service robots to obtain information.Visual perception of environmental information is the main technology for friendly interaction between the service robot and the user.The object detection and target tracking are the key links to realize the intelligence of the service robot.In recent years,object detection and tracking technology have become an important topic in the related field of the robot.In the face of the complex environment,humans can quickly screen out important information and respond with this information.However,robots still not possible to quickly and accurately detect and track objects under large-scale data.Therefore,it is of great significance to improve the accuracy and real-time performance of object detection and tracking by optimizing the robot's mechanism for extracting information.In this paper,we use the inherent reasoning strategy of reinforcement learning to extract information,and study object detection and tracking technology based on Markov decision process(MDP)model.The main contents include:(1)Aiming at the problem of regional extraction for robot detecting objects in the home environments,a Tree-DDQN based on the attention-action strategy is proposed to extract the candidate region which combines the Double DQN(DDQN)with the tree structure.First,the DDQN method is used to select the best action of the current state and obtain the right candidate region with a few actions executed.According to the state obtained after executing the selected action,repeat the above process to create multiple "best" paths of the hierarchical tree structure.The best candidate region is selected by using non-maximum suppression on candidate regions that meet the conditions.Experimental results show that the proposed method based on Tree-DDQN has better detection performance than other methods in most cases.Our method can achieve better in object detection.(2)Aiming at the problem of multi-target tracking,a moving target tracking method is proposed,which is based on corner enhancement and MDP.The paper adopts the MDP model for tracking,which the multi-target tracking problem is converted to a strategic problem based on MDP.A MDP model represents the life cycle of a target,and multiple targets are represented by multiple MDP models.And the paper uses the strong corner points generated by the Shi-Tomasi method instead of the equal distance so that the feature points are more stable in the process of tracking.The learning of similarity function in data association is equivalent to the learning of MDP strategy,which is trained by the reinforcement learning method.The experimental results show that the MDP target tracking algorithm based on cornerenhancement has good tracking performance and capacity of resisting disturbance.(3)In order to study and analyze the object detection and target tracking methods mentioned above,the object detection and tracking system based on the MDP model is designed and implemented.The software system is divided into two parts: object detection and target tracking.Each part is designed and implemented from the training stage and the testing stage respectively.The effectiveness of the algorithm and the robustness of the system are verified on the public data sets and the actual data collected.Based on the above,the service robot can detect the object and track the motion trajectories of the human in the family environment,which improves the visual perception ability of the robot so as to better serve the human.