Traffic Sign Recognition: Benchmark of credal object association algorithms

Static and dynamic objects detection and tracking is a classic but still open problem in Intelligent Transportation Systems. Initially formalized in the Bayesian framework, new methods using belief functions have recently emerged. Most of them have been essentially validated in simulations. This paper proposes an association and tracking framework devoted to Traffic Sign Recognition in a mono-sensor context. Potential signs are detected in the camera images. A credal association between new observations and already known objects is performed. Associated objects are tracked over time and in the image space using Kalman Filtering. This global tracking system has been used to evaluate in real-time on large datasets several state-of-the-art credal association methods. The main evaluation criteria is their capability to reduce false detections by keeping a high traffic sign detection rate.