Human Detection Algorithm Based On Weighted Deformable Part Model

Pedestrian detection technology is a hot but difficult topic in recent years. Firstly, the application of pedestrian detection is very extensive, especially in human-computer interaction, intelligent surveillances and mobile smart devices. Secondly, pedestrian detection is a difficult problem in detection area. Due to human pose articulation, variation in human shapes and appearances, especially occlusion between human and objects, these are serious decrease the accuracy of pedestrian detection. Occlusion problems and pose problems have seriously restricted the smart mobile devices and smart cameras on user experience.Deformable part model(DPM) is one of the most popular pedestrian detection algorithms in recent years. The model employs Histograms of Orients Gradients(HOG) feature and Support Vector Machine(SVM) classifiers. It is reviewed by a coarse root filter which covers the whole objects and several high resolution part filters which cover small part of the objects. Through the score of the root filter and several part filters, we will estimate whether the sliding window contains a human. This coarse-to-fine detection algorithm could achieve a better result. Taking into account that the importance of different part filters is different, we propose a weighted deformable part model based on DPM. Meanwhile, By constructing the relationship between the root filter and the segmented sliding window, we are able to find occluded areas and unoccluded areas. Reasonable adjusting the weight will reduce the inference of occluded area.In this thesis, we generalize and extend the above methods to DPM. Specifically, the contribution of the thesis is two-fold: We propose a method of improving the score of root and part filters by strengthening the weights of key part and reducing the occluded one. Through occlusion inference on sliding window classification results, we propose an approach to adjust the weight of part filters to reduce the negative influence of the occluded area. It has been demonstrated our approach outperforms the state-of-the-art of DPM in INRIA dataset.