Neural Networks for Real-Time, Probabilistic Obstacle Detection

Recent research suggests intrinsically safe robots, such as through soft limbs or artificial skins, to enable close-quarter human-robot collaboration. Intrinsically safe robots allow for risk-minimized instead of collision-free path planning. Risk-minimized path planning can integrate non-binary knowledge—including obstacle probabilities, robot speed, or data age—into the choice of a robot path. In this contribution, we propose a novel approach to probabilistic obstacle detection on color images that is specifically suited for use in real-time risk-minimized path planning. Our approach enhances an existing neural network for object detection by incorporating spatial coherence via a second neural network and an optimization step inspired by simulated annealing. Finally, a bias towards false-positive obstacle detection allows us to avoid the Sleeping Person Problem for online learning. In our experiments, we show that a GPGPU implementation of our approach can process Full HD images at a soft real-time rate of 15 Hz. We conclude that our probabilistic obstacle detection is fit for use in real-time risk-minimized path planning.