Research On Dynamic SLAM Method Based On Image Instance Segmentation

As the country pays more attention to the field of artificial intelligence,mobile robot technology has ushered in significant development in China,and its application in various industries has become more and more extensive,such as unmanned delivery vehicles for express delivery,police robots for street patrol And intelligent service robots in the government hall and so on.With the needs of social development,people have higher and higher demands on the intelligentization of mobile robots.To realize intelligence,autonomous positioning of robots is the key.Visual SLAM technology can help robots achieve autonomous positioning in unknown environments.After decades of development,the SLAM technology has made great progress and the positioning accuracy is getting higher and higher,but the more mature SLAM algorithms currently assume that the environment is static,that is,the positioning accuracy is relatively high in a static environment.When a moving person or a moving object appears,the positioning accuracy will be very low.Therefore,handling dynamic objects in the environment becomes the key to improving the autonomous positioning accuracy of mobile robots.Aiming at the problem that the current semantic segmentation network is difficult to meet the real-time requirements of SLAM operation,this paper designs an instance segmentation network with high real-time performance,which not only ensures the segmentation accuracy,but also meets the real-time requirements of the SLAM system.At the same time,in view of the low positioning accuracy and poor robustness of the current traditional SLAM algorithm in dynamic scenes,this paper improves the existing classic ORB-SLAM2 algorithm by combining image instance segmentation technology and multi-view geometry technology to enhance its motion in dynamic environments.The processing power of the target improves the positioning accuracy of the algorithm.The specific work content of this article is as follows:Firstly,for the problem that most of the current semantic segmentation networks are difficult to meet the real-time requirements of the visual SLAM system,this paper uses the YOLACT instance segmentation network as the basis,and uses the traditional non-maximum suppression(NMS)method to suppress the time-consuming repeated target detection frame.To optimize the problem,a parallel running NMS method is proposed.Compared with the traditional serial running mode,the YOLACT network segmentation accuracy is guaranteed,and its running speed is increased.Finally,the public data set is compared with the currently popular Mask R-CNN instance segmentation network.The results show that the improved instance segmentation network proposed in this paper has better comprehensive performance and better real-time performance,which meets the requirements of SLAM for real-time performance.Secondly,in view of the problem that the dynamic object in the environment causes the SLAM positioning accuracy to decrease,this paper designs a dynamic object segmentation algorithm that combines instance segmentation technology and multi-view geometry.By combining the optical flow method and the multi-view geometry method,the dynamic feature points in the image are tracked and detected,and then the dynamic point detection result and the instance segmentation thread segmentation result are combined to determine the dynamic object area in the image and segment it.Then,in view of the low positioning accuracy of the current traditional SLAM algorithm in dynamic scenes,this paper proposes an improved SLAM algorithm based on dynamic object segmentation.In this paper,the current classic ORBSLAM2 algorithm is selected as the basic framework for improvement,and the dynamic object segmentation algorithm designed in this article is integrated into the visual odometer part to realize the detection and segmentation task of dynamic targets in the environment,and then use images without interference of dynamic objects Performing feature matching and pose estimation improves the robustness of the algorithm in dynamic scenes.Finally,the public data set is used to verify the effectiveness of the dynamic object segmentation algorithm proposed in this paper,and the pose estimation experiment of the improved ORB-SLAM2 algorithm and the algorithm before the improvement is carried out.The results show that the dynamic target segmentation algorithm proposed in this paper is effective At the same time,the improved ORB-SLAM2 algorithm proposed in this paper can achieve higher positioning accuracy and smaller errors in both high dynamic environment and low dynamic environment compared to the previous algorithm,and it has better robustness in dynamic scenarios.It also meets real-time requirements.