High Density People Estimation in Video Surveillance

Pedestrian detection is a challenging problem in computer vision due to difficulties such as occlusion and diversity of pedestrian visual features. The complexity of the problem increases as the diversity of the pedestrians in the scene being analyzed increases. Different approaches have been proposed. Among these approaches, the ones that involve supervised techniques for head detection have shown robust results and better accuracy. In this paper, we provide a comparison of the accuracy and performance of cascaded classifiers based on two different feature extractors, notably, the Haar-Like and the Local Binary Pattern (LBP) features extractors. It has been found that increasing the diversity and size of the training dataset leads to improvements in the resulting detectors. Therefore, four different training datasets have been constructed using standard training sets as well as our own created images for crowded scenes. Moreover, we introduce a tuning scheme for the parameters of the detector. The results obtained exceeded 80% for precision on high density images and more than 90% on lower density images.