Robust Tracking Approach for Dealing with Classification Uncertainty

Abstract

Every year about 1.25 million people die as a result of road traffic accidents. Besides the traffic on the road increases every day, including the environmental impact due to the corresponding traffic emissions. Autonomous driving could be a unique opportunity to increase these traffic safety, traffic flow efficiency and to reduce emissions in future. In order to operate reliably and accurately, autonomous driving vehicles and autonomous features require an accurate perception of the infrastructure and other road users in the surrounding. Multi-object tracking is the process concerned with the estimation of the states of the objects in the environment, given the noisy measurements from the sensors. Besides the estimation of the states, it is also necessary to estimate the classification of the objects e.g. for a correct situation analysis and path planning. Classifiers are used to detect and classify objects from image frames, however the classification is sometimes incorrect or uncertain. This results in a decrease of tracking accuracy and an incorrect classification in these classification uncertain conditions. The contribution in this work is, by keeping the classification uncertainty in the tracking approach and using it in all steps, to jointly improve the tracking accuracy as well as the classification.