NEW APPROACH FOR CYCLIST DETECTION BASED ON PROBALISTIC FLOW MAPS AND SPECTRL ANALYSIS

摘要

A method to detect and classify cyclists was introduced. The classification is done by spatial and spectral motion analysis. The benefit of the method is the parallelization of the tracking and the classification task. The classification itself is the result of the motion analysis during the tracking steps. In contrast to related works, there is no use of geometrical analysis like Hough transformations, especially no detection of bicycle wheels. Instead, the system is capable to detect the typical pedal movement of a cyclist from different positions. Therefore a detection cascade is shown which is, at its first step, oriented on well-known pedestrian detection. Histogram of oriented gradients is combined with a support vector machine to detect pedestrians. The cyclist features like pedal movement are evaluated in a second step in the spatial and temporal domains. A stable pattern of the pedal movement is generated. The feature movement is evaluated during a predefined period using the cosine transform. With typical pedal movement the spectrum can be used classify the pedestrian as cyclist. Test and results are discussed and presented.