LiDAR-based Pedestrian Tracking Adapting to Sparse Point Cloud Utilizing Interacting Multiple Model

The vehicle navigating through narrow and crowded environments requires detailed shape information of the surrounding pedestrians for collision avoidance. While Light Detection and Ranging (LiDAR) is highly effective at measuring the position of objects, its performance diminishes as the distance between the LiDAR and the objects increases. The number of data points acquired decreases, leading to a less informative reconstruction of the object’s pose and shape. To address this issue, this study proposes a pedestrian tracking method that involves constructing multiple pedestrian models and estimating the appropriate model parameters from likelihoods using the Interacting Multiple Model, based on the number of point clouds. We prepare three models: ellipse, bounding box, and point cloud center of gravity models. The ellipse and bounding box models estimate pose and size, while the point cloud center of gravity model estimates pose. The elliptical model uses Random Sample Consensus to determine model parameters that suppress arm swing and body sway during walking. Through experimental validation, this method effectively demonstrated its ability to continuously track pedes- trians, including those with only a few acquired data points from a pedestrian located far from the LiDAR, while accurately estimating their pose and size.