An Innovative multi-sensor data fusion method based on thrice deeply-fusion architecture for multi-object tracking

Abstract

For the purpose of track maintenance in a cluttered environment with some false alarms, false dismissals and measurement disturbances, this paper presents a new multi-object tracking procedure with uncertainty of the source of the measurement data returned from multi-sensor. The tracking method is a thrice deeply-fusion approach constructed by a primary fusion based on an improved probabilistic data association filter (IPDAF), a secondary fusion with historical motion trajectories, and a thrice fusion with road-markings. It incorporates the existence probabilities of the individual tracks and the variable number of objects based on the Bayesian estimation theory, which can improve the tracking performances effectively in an environment with high clutter density. A binary 2-D assignment is adopted for the optimal data association, which is established as a nonlinear optimization problem. In the motion modeling, it introduces multiple measurement models for different sensors into the method. Then, the estimation could be performed with greater reliability. The computational efficiency is satisfying and it can be used for real-time application, which is verified by two real test scenarios.