Closed-loop detection of mobile robot based on quaternion convolutional neural network

To improve the accuracy of closed-loop detection for mobile robot in complex environments and reduce the cumulative error of visual odometry, a closed-loop detection method based on quaternion convolutional neural network is proposed. Firstly, the superpixel segmentation is used to extract multi-scale landmarks and improve the appearance invariance and viewpoint invariance of image description. Then, the convolutional layer of convolutional neural network is extended to the quaternion convolutional layer, which can effectively increase the correlation of red, green and blue channels, extract the deep information of color images, and reflect the integrity of color images. Finally, in image similarity measurement, not only the distance of road sign is calculated, but also the shape and spatial distribution information of landmarks are considered to improve the accuracy of similarity measurement. The effectiveness of the proposed method is experimentally verified on the public datasets of Carnegie Mellon university and the University of Queensland. Experimental results show that, compared with traditional closed-loop detection algorithm, the proposed algorithm not only guarantees a high recall rate, but also improves the accuracy of closed-loop detection, achieving an average accuracy of 87.64% and 90.12% respectively on the data set, significantly improving the precision of closed-loop detection of mobile robots in complex environments such as such as illumination variations, view variations, and seasonal variations. © 2019, Editorial Department of Journal of Chinese Inertial Technology. All right reserved.