Research On The Closed-loop Detection Method Of Vision SLAM Based On Deep Learning

Visual SLAM which uses camera as sensor is a branch of SLAM technology,and is an important technology for autonomous navigation and mapping of robots in unknown areas.The vision SLAM system will output the corresponding map after the motion estimation between the front frames and the optimization of the back end.Due to the error accumulation caused by the long-term movement in a large range,closed-loop detection is needed to verify whether the robot has reached the scene in real time,so as to correct the drawing in time.The traditional closed-loop detection method is mostly based on the visual word bag model,and the scene features extracted by the traditional method are all set artificially,so the accuracy of the traditional method is not high and the time is long.In order to satisfy the demand of accuracy and real-time of closed-loop detection,the following researches are carried out in this paper.Aiming at the problem of low accuracy of traditional algorithm to extract features,this paper uses the improved lightweight convolutional neural network MobileNetv3 model to input the image to be tested into MobileNetv3 after pre-training with ImageNet to extract features,construct the feature vector matrix,and then use cosine distance to match the image similarity,so as to complete the closed-loop detection.Experimental results show that the accuracy of the closed-loop detection algorithm is significantly higher than that of traditional algorithms and existing algorithms based on other convolutional neural networks.Aiming at the time-consuming problem of traditional algorithms,three classical dimensionality reduction methods are used in this paper for longitudinal comparison,and it is found that KPCA has a good effect.Therefore,this paper proposed the Mobilenetv3-KPCA closed-loop detection algorithm using Kernel Principal component analysis(KPCA),in the case of ensuring high accuracy,we reduce the high-dimensional image feature matrix of MobileNetv3 output from 1,000 to 500 dimensions,the experiment proved that the algorithm after dimension reduction time performance improved about 40%.In addition,transverse comparison,which includes comprehensive comparison of the proposed algorithm and the six closed-loop detection algorithms in good performance of existing literature,again demonstrates that the proposed MobileNetv3-KPCA algorithm has higher accuracy and better real-time performance.To sum up,the MobileNetv3-KPCA algorithm proposed in this paper performs well and can satisfy the demand of visual SLAM system on the accuracy and real-time of closed-loop detection.