Laser-based target tracking using principal component descriptors

The reliable detection and tracking of general objects is required by many field robotics applications, where autonomous agents need to navigate between and interact with dynamic targets in unstructured environments. This paper presents an approach to the detection and tracking of both moving and stationary objects in a forward-facing laser scan. Traditional approaches use geometric primitives to detect and model specific targets. A more general target descriptor taking object location and size into account is presented here, using principal component analysis to extract these features. Kalman filtering using a white noise acceleration model is implemented to track objects, with extensions to the target motion model provided in order to account for laser scanner motion. Results presented show that the proposed system tracks targets effectively over a wide range of challenging situations.