Labeled Multi-Bernoulli Filter for Distributed multi-target Tracking with Detection and Class Measurements

Advanced machine vision and deep learning models are now widely used as virtual sensors in various applications to detect and classify objects in image measurements. Typically, these virtual sensors output measurements as a set of class probabilities for each detected object within the sensor’s field of view. However, integrating this type of data into multi-target tracking systems, traditionally based on point measurement detections, presents some challenges. This paper proposes a solution by introducing a labeled multi-Bernoulli filter formulated to handle detections with class measurements from virtual sensors for multi-target tracking. Furthermore, we explore the application of this filter for multi-sensor multi-target tracking within a distributed sensor network. Through numerical experiments involving vehicles equipped with cameras and deep learning classification modules navigating complex roads, we demonstrate that incorporating class information into the tracking process improves tracking accuracy. This improvement is further enhanced when vehicles share and fuse their information in a distributed manner. Our findings highlight the benefits of integrating class probability data into multi-target tracking filters, offering substantial improvements in the accuracy of automated monitoring systems in dynamic and complex environments.