A Research On Pedestrian Detection Method Based On Multi-Models Matching

Pedestrian detection is one of the key research directions in computer vision, with several potential applications in many fields, including auto-auxiliary driving system, military reconnaissance, and intelligent video surveillance system. In recent years, many effective pedestrian detection algorithms had been proposed, most of them follow a dense multi-scale sliding-window paradigm that binary classification detection windows with the trained model. Therefore, both the feature extraction algorithm and the classifier are critical factors for obtaining good performance.Pedestrian detection is a challenging task owing to their highly variable appearance and complex backgrounds. Here we concentrate on the classifier, linear SVM(Support Vector Machine) was adopted as a test case, and two efficient pedestrian detection methods based on multi- models matching are presented to adapt pedestrian appearance change better, summarized as follows:Firstly, we put forward a multi- models matching algorithm based on clustering algorithm. Features which insensitive to color and light were selected, take histogram of oriented gradient(HO G) feature as an example. In the training phase, cluster algorithm was introduced to separate training human samples, so samples with similar appearance, such as posture, view-points, are gathered into a category, and then several models were trained with these different clusters respectively, to improve its adaptation to corresponding human cluster. In the test phase, trained models were matched independently and the repeat detections from different models were combined with linear weighting to obtain the final test result. Different clustering methods(hierarchical clustering, K-Means and FCM) and cluster number(2, 3 and 4) were experimented, Simulation results show that this method has more competitive performance than HOG algorithm on the INRIA and ETH pedestrian dataset.Secondly, a multi- models matching algorithm based on cascade training was proposed. In the training phase, several models were cascade trained with different features, selected by complementary feature selection strategy(CFSS), to improve its adaptationto hard samples, HOG feature of RGB and LUV color channels and cell-structured local binary patterns(LBP) feature were selected. In the test phase, trained models were matched independently and the repeat detections from different models were combined with linear weighting to obtain the final test result. Simulation results show that, our method has more competitive performance than four methods of VJ, HOG, Hog Lbp, and Multi Ftr on the INRIA pedestrian dataset.