Depth Image-Based Obstacle Avoidance for an In-Door Patrol Robot

Image-based obstacle avoidance has been studied for decades. One weak point of image-based approaches is that the performance usually depends on the lighting condition. That is, the performance can be very poor in dark environments. In this research, we investigate the possibility of the depth image-based approach for full-time indoor patrolling. As the first step, we consider a 3-class problem. Each depth image is classified as “danger” if some obstacle is too close, as “notice” if the obstacle is close, and as “normal” if there is no obstacle in the vicinity. The label of each depth image is defined based on the RGB image captured at the same time, and an AlexNet, which is a well-trained convolutional neural network, is retrained via transfer learning, and used for classification. In our primary experiment, we collected 102,776 image data in the Research Quadrangle of the University of Aizu. Test results show that the performance of the depth image-based approach is good during both day and night, and in most cases, it is better than the RGB image-based approach. This result can provide new insights when designing more practical full-time patrol robots.