Detection and Recognition confidences update in a multi-sensor pedestrian tracking system

Fadi FAYAD, Véronique CHERFAOUI.

We present in this paper a method for confidence updating in a multi-sensor pedestrian tracking system. We focus on the computing of detection and recognition confidence indicators according to the object detection and/or recognition probabilities provided by sensor modules. We propose to use the Transferable Belief Model in order to model and combine the sensor module outputs at different times. Since detection and recognition processes are not really independent, we propose to use the cautious rule to combine the belief functions. We propose a track confidences updating algorithm and its interesting behavior is shown on synthetic data.

PDF full paper