A solution for probabilistic inference and tracking of obstacles classification in urban traffic scenarios

Obstacles classification plays an important role in driving assistance systems. Any classification system should accurately distinguish, in real-time, between a set of well-known object classes such as pedestrians, cars and poles and other obstacles. If the object class is determined then the driving assistance system may take the right decision, in case of an imminent impact, in correlation to the vulnerability of the class that object belongs to. An object detection module based on both 2D and 3D information is considered for the obstacles segmentation. Preliminary classification results are obtained, at each image frame, for each detected object. The classification result should be approximately the same for an object that is tracked across frames. We described some methods for accomplishing this issue. First a Bayesian inference is considered for obtaining the class probability of the tracked objects from frame to frame. Then the tracking and filtering of the object´s class is realized by applying a k-NN classification on the previously computed class values over the last few frames. These methods improve the stability and accuracy of tracked objects´ classification across multiple frames.