Evidential data association based on Dezert-Smarandache Theory

Data association has become pertinent task to interpret the perceived environment for mobile robots such as autonomous vehicles. It consists in assigning the sensor detections to the known objects in order to update the obstacles map surrounding the vehicle. Dezert-Smarandache Theory (DSmT) provides a mathematical framework for reasoning with imperfect data like sensor's detections. In DSmT, data are quantified by belief functions and combined by the Proportional Conflict Redistribution rule in order to obtain the fusion of evidences to make a decision. However, this combination rule has an exponential complexity and that is why DSmT is rarely used for real-time applications. This paper proposes a new evidential data association based on DSmT techniques. The proposed approach focuses on the significant pieces of information when combining and removes unreliable and useless information. Consequently, the complexity is reduced without degrading substantially the decision-making. The paper proposes also a new simple decision-making algorithm based on a global optimization procedure. Experimental results obtained on a well-known KITTI dataset show that this new approach reduces significantly the computation time while preserving the association accuracy. Consequently, the new proposed approach makes DSmT framework applicable for real-time applications for autonomous vehicle perception.