3D target tracking and shape reconstruction in clutter using Gaussian process and point completion network

Existing problems for 3D extended target tracking and contour reconstruction under clutter environment are thoroughly investigated in this article. Due to the sparsity of available point cloud data and interference from clutter, the shape completion method, such as Point Completion Network (PCN), cannot reconstruct 3D contour directly. Moreover, the traditional Gaussian process (GP) model suffered from computing overhead and cannot handle irregular non-convex shapes. Here, a two-stage tracking algorithm is proposed, which firstly combines the 3D GP measurement model with a probability data association filter to jointly estimate the kinematics and 3D Gaussian basis points, which contain partial shape information. Afterwards, this method input produced Gaussian basis points to a deep learning-based PCN to reconstruct complete geometry of the contour. The effectiveness of the proposed algorithm is verified by extended target tracking with both simple and complex geometric shapes. The simulation results show that the proposed algorithm obtains accurate kinematic state estimation and produces complete point cloud estimation for 3D contour.