Correlation between features and classifiers for semantic understanding of pedestrian attitudes in traffic scenes

Within the context of a traffic scenario, pedestrians may have several attitudes or perform different actions: wait at the traffic light, cross the street, run for a bus or a taxi, walk or run on the pavement. When performing all these actions, pedestrians have different attitudes: stand, walk, run. We have studied those attitudes and the contexts in which they appear and we have derived some semantic concepts. For each concept we have extracted different features and using the features we have trained classifiers. The novelty of the paper resides in presenting the correlation between semantic concepts that characterize the attitude of a pedestrian, feature type and classification scheme. As features we have used: histograms of oriented gradients and Gabor wavelets. For classification we have experimented adaptive boosting, neural networks and support vector machines. The system uses stereovision for computing the speed and orientation of the objects that appear in a traffic scene. A stereo based pre-classification step is used for determining the speed, orientation and the dimension of the pedestrian pattern.