Autonomous Mobile Robot Visual SLAM Based on Improved CNN Method

Abstract

Although the research on autonomous mobile robot SLAM has received extensive research, the current mobile robot still exists in practical applications: it may move under the condition of disordered and irregular obstacle placement; the shape of the obstacle and the position of the obstacle change; and indoor and outdoor scene switching occurs at different times and other issues. Autonomous mobile robots need to continuously optimize SLAM during motion and obtain real-time information from multiple sensors for real-time identification and rapid response to the surrounding environment. We have improved the CNN-based VSLAM system by replacing the original single convolutional layer with parallelism and reducing the number of model parameters. This approach can reduce the problem of system gradient disappearance and continue to train deeper networks. Finally, a global map is generated with the fully connected layer and passed to the robot's navigation. The experimental results based on RGB-D SLAM Dataset and Benchmark database dataset show that the proposed VSLAM system based on CNN is superior to the traditional CNN VSLAM system in both ATE and RPE indicators.