A Dynamics Observation Model for Moving Object Based on Reinforcement Learning: Experimental Demonstration on Stereo Vision

In this article, a dynamics observation model is proposed to obtain the information of the pose of a moving object and its higher order motion, which is one of the fundamental tasks in the field of target recognition. First, an object data model based on sparse feature line segments is constructed. Second, a dynamics observation model inspired by reinforcement learning is proposed with the learning of the attitude and cost dynamics as the core. In attitude dynamics learning, an optimal policy iteration method is established based on dynamics principle. An asymptotically convergent differentiator is designed in cost dynamics learning, which improves interference immunity and accuracy by the smooth saturation function and the integral of error function. The proposed model enables the real-time observation of the dynamics process of an object, by using only the coordinate information of sparse key points. It can preserve high accuracy and fast convergence even when the key points exhibit sparse distribution. Finally, the effectiveness and superiority of the proposed observation model are verified through numerical simulations and depth camera-based physical experiments.