Moving Direction Predicting with Deep Neural Networks for Mobile Robots

Localization and navigation is a challenge in the robot field. At present, the commonly used methods include GPS/INS integrated navigation and simultaneous localization and mapping (SLAM). Integrated navigation does not need to map the scene, which can adapt to the unknown environment well. But the satellite signal is easy to be jammed, resulting in large positioning error. The SLAM method does not depend on external signals, and has good adaptability to occlusion, especially the indoor environment. But SLAM needs to establish a map in advance. It is very difficult and high costly for robots to automatically establish a new high-precision map in a strange environment, and the map has no directionality without manual labeling. Aiming at the above problems, this paper proposes an image navigation method based on a deep neural network to determine the motion direction of the robots. In contrast to the traditional localization-planning navigation methods, the proposed method directly models the relevance between images and final navigation results, which simplifies the complex navigation problem into a pattern classification problem. In order to verify the effectiveness of our method, we collected a set of campus data and designed experiments. The results show that the method can effectively indicate the moving direction for robot navigation.