Research On Dynamic Scene Visual SLAM Based On Multi-sensor Fusion

With the vigorous development of intelligent robot technology,the application of simultaneous localization and mapping technology is becoming more and more extensive.The dynamic factors in the outdoor open environment greatly affect the accuracy and stability of SLAM algorithm.Fusion of multiple sensor input data can improve the accuracy of SLAM algorithm,but the data association between sensors also brings new challenges to the deployment and implementation of SL AM algorithm.Based on the above reasons,this paper studies the vision SLAM algorithm based on multi-sensor fusion in dynamic scenes,and proposes a SLAM framework that integrates lidar,camera and Inertial Measurement Unit sensors.This framework can not only maintain certain stability in the environment with many dynamic scenes,but also automatically correct the initial external parameters inaccuracy between sensors.The main work of this paper is as follows:(1)Aiming at the problem that the current neural network model used for vision SLAM is not effective in detecting and segmenting small objects,a method of image semantic segmentation based on neural network is proposed.Its core is the multi-class Focal Loss function used for training neural network.The loss function can not only statically weight the loss value of each category according to the frequency of each category of pixels in the data set at the initialization time,but also dynamically weight the loss value of each category according to the training effect of each round during the training process.The experiment shows that the segmentation ability of the model for small class objects is improved after using the loss function for training.(2)Aiming at the problem that the initial external parameter calibration error between sensors or the external parameter distortion caused by motion affects the data fusion effect of the front end sensor of SLAM,a real-time external parameter correction module for the front end of SLAM is proposed.This module extracts semantic features based on the laser radar point cloud and camera image respectively,and carries out matching construction constraints for optimization,and corrects the external parameters between sensors according to the optimization results.The experiment proves that this module can make the SLAM system overcome the initialization failure caused by external parameter errors to a certain extent,and ensure that the tracking accuracy will not decline due to the accumulation of wrong external parameters during a long operation.(3)Aiming at several kinds of dynamic objects commonly seen in dynamic scenes,a dynamic feature filtering module is proposed.Use the depth learning algorithm based on multi-class Focal Loss proposed in this paper to segment the camera image,classify the features extracted from the front end of SLAM according to the projection relationship,screen out unreliable dynamic features,and improve the accuracy of the algorithm.In summary,the proposed dynamic scene vision SLAM algorithm framework for multi-sensor fusion has been validated based on the KITTI Odometry dataset,and compared with the R3Live algorithm and LVI-SLAM algorithm.Experimental results show that the proposed dynamic scene vision SLAM algorithm is better than that of the other two algorithms in positioning error and runtime in most scenarios.