Visual Simultaneous Localization and Mapping Algorithm Based on Convolutional Neural Network to Optimize Loop Detection

Traditional visual SLAM (simultaneous localization and mapping) without loop detection may lead to error accumulation. Even if there exits loop detection, it is unable to be applied to the lightweight applications because of its low accuracy and efficiency. Thus, a visual SLAM with loop detection optimization is studied. In the front-end estimation, ORB (oriented fast and rotated brief) feature points were abstracted and matched. PnP (perspective-n-point) was solved for the successful matched point to estimate the camera motion and screen out the key frame images. In the back-end optimization, SqueezeNet convolution neural network (CNN) was used to extract the feature vectors. The cosine similarities were calculated to determine whether there were loops or not. If there was a loop, the corresponding constraint was added to the posture graph. Then the global posture was optimized by using the graph optimization theory. Finally, tests and comparisons were conducted on the data sets produced by our research group and the public data sets of TUM. The results show that the proposed algorithm can detect loops successfully and add constraints to global trajectory optimization compared with the non-loop detection algorithm. Compared with the traditional word bag method, the recall rate of this method can be increased by 21% and the calculation time can be reduced by 74% under the same loop detection accuracy. Compared with RGB-D SLAM algorithm, the error of this method can be reduced by 29%.