Vision-Based Obstacle Avoidance Using Deep Learning

This paper describes a vision-based obstacle avoidance strategy using Deep Learning for Autonomous Underwater Vehicles (AUVs) equipped with a simple colored monocular camera. For each input image, our method uses a deep neural network to compute a transmission map that can be understood as a relative depth map. The transmission map is estimated for each patch of the image to determine the obstacles nearby. With this map we are able to identify the most appropriate Region of Interest (RoI) and to find a direction of escape. This direction allows the robot to avoid obstacles by performing a control action. We evaluate our approach in two underwater video sequences. The results show the approach is able to successful find a RoI that avoids coral reefs, fish, the seafloor and other object present in the scene.