Elliptical Model-Based Robust Pedestrian Pose Tracking by LiDAR via Bi-Level Moving Horizon Random Sample Consensus

Recent advancements in autonomous vehicle research have heightened the need for precise pedestrian pose tracking to ensure safe navigation. Utilizing LiDAR sensors, conventional bounding box (BB) methods are generally used but are sensitive to limb motion. Previous moving horizon-based maximum likelihood estimation sample consensus (MHESAC) study addressed this robustness limitation but suffered from extensive point cloud (PC) sampling, limiting real-time applications. This paper proposes an elliptical pedestrian model tracking within a sequential bi-level MHESAC (B-MHESAC) framework that reduces PC sampling while preserving high accuracy. In the first level, elliptical models are locally fitted to the pedestrian torso for each time step using random sampling consensus (RANSAC). In the second level, sequences of local RANSAC models over the moving horizon are sampled and evaluated using maximum likelihood estimation sample consensus (MLESAC) to improve model selection. For a sufficiently large outlier ratio, B-MHESAC requires less trials than MHESAC when considering multiple horizons and samples. Moreover, at long horizon lengths, B-MHESAC required trial's exponential growth is at least reduced to the fifth root of MHESAC, allowing real-time estimations. The proposed method was evaluated in an indoor single pedestrian tracking experiment. Experimental results demonstrate that B-MHESAC is more robust than conventional RANSAC in model fitting. Additionally, the elliptical model-based B-MHESAC's tracking performance is competitive with the MHESAC method's and superior to other Kalman filter-based BB and circle model methods.