Fast Object Classification for Autonomous Driving Using Shape and Motion Information Applying the Dempster-Shafer Theory

The classification of moving objects is one of the crucial tasks to enable autonomous driving in urban environments. Even in simple traffic scenarios we already observe a multitude of dynamic objects such as pedestrians, bicyclists or cars. The challenge is to accurately determine the correct class of each object under all circumstances. This problem gets even harder if an object transitions from one class to another, e. g., a person that first pushes a bicycle and then starts to ride it. The question we are facing here is: How can we represent this situation in terms of classification?We propose a universal way of describing the characteristics of the classes by designing specialized belief functions to model object features. Given the belief functions, so-called “mass values” can be derived for each class. Finally, using the Dempster-Shafer theory, a “belief” is assigned to the power set of the classes. The approach provides a direct correlation to the features and is, therefore, able to provide a comprehensible explanation for the result.We evaluate the proposed approach given data gathered with a LiDAR sensor. We present its classification performance, show its effectiveness and discuss advantages and disadvantages.