Predicting collisions: time-to-contact forecasting based on probabilistic segmentation and system identification

The Time-to-contact (TTC) estimate is mainly used in robotics navigation, in order to detect potential danger with obstacles in the environment. A key aspect in a robotic system is to perform its tasks promptly. Several approaches have been proposed to estimate reliable TTC in order to avoid collisions in real-time; nevertheless they are time consuming due to a calculation of scene characteristics in every frame. This paper presents an approach to estimate TTC using monocular vision based on the size change of the obstacles over time (); therefore, the robotic system may react promptly to its environment. Our approach collects information from few data of an obstacle, then the behavior of the movement is found through an online recursive modeling process, and finally, a forecasting of the upcoming positions is computed. We segment the obstacles using probabilistic hidden Markov chains. Our proposal is compared to a classical color segmentation approach using two real image sequences, each sequence is composed of 210 frames. Our results show that our proposal obtained smoother segmentations than a traditional color-based approach.