Loop Closure Detection for Visual SLAM Systems Based on Convolutional Netural Network

In this paper, the loop closure detection technology is studied. Aiming at the problem that the use of artificially marked feature points in the traditional visual SLAM algorithm leads to a significant decrease in the accuracy of the loop detection algorithm in a complex environment and an environment with obvious lighting changes, this paper proposes a loop closure detection algorithm based on deep learning. Firstly, the YOLOv4 model with optimized loss function is used to detect the target in the images collected by the camera. Then, the Locality Sensitive Hash function is used to reduce the dimension of high-dimensional data, and the loop is determined according to the cosine distance. Finally, the simulation results show that the algorithm can reduce the cumulative error of the robot, obtain the global consistency map, and achieve better results in real-time and accuracy.