Information fusion to detect and classify pedestrians using invariant features

A novel approach to detect pedestrians and to classify them according to their moving direction and relative speed is presented in this paper. This work focuses on the recognition of pedestrian lateral movements, namely: walking and running in both directions, as well as no movement. The perception of the environment is performed through a lidar sensor and an infrared camera. Both sensor signals are fused to determine regions of interest in the video data. The classification of these regions is based on the extraction of 2D translation invariant features, which are constructed by integrating over the transformation group. Special polynomial kernel functions are defined in order to obtain a good separability between the classes. Support vector machine classifiers are used in different configurations to classify the invariants. The proposed approach was evaluated offline considering fixed sensors. Results obtained based on real traffic scenes demonstrate very good detection and classification rates.