Research On Global Localization And Moving Objects Detection For Mobile Robot

Mobile robot is a kind of autonomous robot system that can sense the environmental information and estimate its own state through the equipped sensors.It can realize goal-oriented autonomous movement in a complex and dynamic environment.Generally,due to the complexity of the environment and the limitations of the sensor's perception ability,the information obtained by the mobile robot usually is uncertain and incomplete.However,understanding the environment accurately is an important precondition for accomplishing its given tasks.Among the tasks,the detection of the moving object based on the incomplete information obtained is the key to improve the robustness and stability of the autonomous robot in a dynamic environment.The perception of moving object for mobile robots can be mainly divided into two sub-problems:self-state estimation and moving objects detection.Based on the established environment map and observed environmental information,the process of the robot pose estimating can be considered as global localization.Due to the low cost of camera sensors and the abundant environmental information provided,visual based global localization is the mainstream method now.There are two main problems in the visual based global localization methods.First,visual information can be greatly affected by ambient lighting,occlusion,and viewpoint.It is difficult to extract robust visual features and obtain stable description and recognition of scenes.Second,uncertainty and incompleteness are the essential attributes of visual information,which requires an effective information fusion mechanism to achieve accurate global pose estimation.To tackle these problems,in our work,the convolutional neural network is used to express the environmental information,and a more robust scene recognition is realized under the condition that the apparent appearance of the environment changes greatly.Furthermore,a visual global localization algorithm based on the particle filter is further proposed.Combined with the context information of the image sequence,the ambiguity of global localization based on a single image are reduced.Experimental results show that the algorithm effectively improves the recall of global localization.Under the premise that the global pose of the mobile robot is known,the moving object detection problem for mobile robot during the motion can be approximately transformed into a problem of moving objects detection with a static background.Aiming at improving the detection accuracy and recall,a novel moving objects detection method based on 2.5D grids is proposed.The method construct a stack of grids iteratively based on the background obtained at the last time.As the background model can be updated in real time,the accuracy of moving objects detection and tracking robustness are improved.Experimental results show the improvement of our algorithm in terms of the accuracy and recall of detection and tracking robustness.