Real-time human detection in urban scenes: Local descriptors and classifiers selection with AdaBoost-like algorithms

This paper deals with the study of various implementations of the AdaBoost algorithm in order to address the issue of real-time pedestrian detection in images. We use gradient-based local descriptors and we combine them to form strong classifiers organized in a cascaded detector. We compare the original AdaBoost algorithm with two other boosting algorithms we developed. One optimizes the use of each selected descriptor to minimize the operations done in the image (method 1), leading to an acceleration of the detection process without any loss in detection performances. The second algorithm (method 2) improves the selection of the descriptors by associating to each of them a more powerful weak-learner - a decision tree built from the components of the whole descriptor - and by evaluating them locally. We compare the results of these three learning algorithms on a reference database of color images and we then introduce our preliminary results on the adaptation of this detector on infrared vision. Our methods give better detection rates and faster processing than the original boosting algorithm and also provide interesting results for further studies.