Multi-feature real time pedestrian detection from dense stereo SORT-SGM reconstructed urban traffic scenarios

In this paper, a real-time system for pedestrian detection in traffic scenes is proposed. It takes the advantage of having a pair of stereo video-cameras for acquiring the image frames and uses a sub-pixel level optimized semi-global matching (SORT-SGM) based stereo reconstruction for computing the dense 3D points map with high accuracy. A multiple paradigm detection module considering 2D, 3D and optical flow information is used for segmenting the candidate obstacles from the scene background. Novel features like texture dissimilarity, humans´ body specific features, distance related measures and speed are introduced and combined in a feature vector with traditional features like HoG score, template matching contour score and dimensions. A random forest (RF) classifier is trained and then applied in each frame for distinguishing the pedestrians from other obstacles based on the feature vector. A k-NN algorithm on the classification results over the last frames is applied for improving the accuracy and stability of the tracked obstacles. Finally, two comparisons are made: first between the classification results obtained by using the new SORT-SGM and the older local matching approach for stereo reconstruction and the second by comparing the different features RF classification results with other classifiers´ results.