Learning an Image-based Obstacle Detector With Automatic Acquisition of Training Data

Other than element of y(t) is 1. detecting whether an obstacle is present, the system also estimates its position. Instead of attempting to reconstruct the 3D structure of the environment in front of the robot, we follow a conceptually simpler and computationally lighter approach which considers each frame independently and does not rely on a sophisticated computer vision pipeline. As humans, when we observe a single picture, we can instinctively infer where an obstacle is present and which areas are free; this is because we have a prior expectation (learned from experience) on the appearance of free space and obstacles, not because we performed a multi-view 3D reconstruction of the scene. In order to achieve a similar goal, our approach works Figure 1: Images paired with corresponding proximity sensor by acquiring training datasets on a robot that is equipped data (groundtruth) with both a camera and a number of proximity sensors that can detect obstacles in the same area imaged by the camera, and thus produce a ground truth.