Evaluation of State-Dependent Pedestrian Tracking Based on Finite Sets

This paper addresses the use of multiobject filters based on finite set statistics with a special focus on sensor characteristics for the use in distributed indoor pedestrian tracking with multiple sensors. For this purpose, a sensing framework is presented consisting of Lidar sensors together with depth cameras. To make use of previously existing knowledge about the measuring process, an adaptive sensor model is presented with a focus on state-dependent modeling of the sensor characteristics. The model incorporates changes in the probability of detection due to the distance between the object and the sensor, occlusions, and the sensor location-dependent environment. The model is incorporated into three approximations of a Bayes optimal multiobject filter, namely, the probability hypothesis density filter, the cardinalized probability hypothesis filter, and the cardinality balanced multitarget multi-Bernoulli filter. Using labeled sequential Monte Carlo implementations, the filters are evaluated with and without the proposed adaptive sensor models for partly simulated and real data. The evaluation was done using ground truth data obtained by a marker-based motion capture system. It is shown that the adaptive model achieves superior tracking results in simulation and real experiments for all approximations with the cardinalized probability hypothesis filter producing the best results.