Research Of Pedestrian Detection Based On DPM And Pedestrian Feature Extraction

Pedestrian detection is one of the most focused research in computer vision field, which has been widely used in various fields in real life. This paper mainly aims at the pedestrian detection method and pedestrian feature extraction algorithmDeformable part model is one of the most used methods for pedestrian detection. This paper first analyzed the deformable part model, and improved its performance according to its characteristics and model structure. DPM uses a single feature for which leads to a lack of description for the characteristics. To solve this, this paper proposes a multi-features fusion method, which selecting complementary texture feature and color feature for fusing. Using the rotation invariant and uniform local binary pattern as texture feature. For color feature, this paper proposes a new color self-similar feature, using color names features instead of the HSV color histogram in CSS and simplifies the similarity calculation method. For the object occlusion problem, the DPM model structure is improved. This paper puts forward a simple weighted part model, which is built after the DPM's training stage. The weight of each part filter can be got after retesting on the test dataset and calculating according to the threshold. The new weighted part model can well detect the pedestrian image in which occlusion is existed.Extraction of the pedestrian related features can help identify person information effectively. This paper has carried on the extraction of pedestrian related features based on the color and texture of people's clothes. At first the image enhancement methodis used to get a better image. Then the grabcut algorithm and gradient threshold value can be used to get the segmentation model of person. Based on the segmentation model the color and texture features can be extracted, which are then used for identifying. Using color names as color features. Using the improved completed LBP features combine with rotation invariant HOG features as the texture features.Then the SAE classifier is used to classify.The grabcut segmentation in last method has some disadvantages. For comparing,this paper using the parts' locations estimation based on DPM and the superpixel regions labeling method. The part model's structure is changing in the DPM model,joining the co-occurrence model. The gPb-OWT-UCM segmentation algorithm is used to divide the image into superpixel region. The labels of superpixel regions are estimated using CRF model and the parts' position constraint. Then the segmentation of an object can be get according to the merging of superpixel regions with same labels. Using this classification model for extraction of pedestrians related feature, the accuracy can be improved.At last, this paper built a pedestrian detection system. It can carry on the pedestrian detection and the pedestrian related features' extraction. The detected pedestrians and its related information will be stored in the database and can be searched using the key input.