The Research Of Pose Varied Pedestrian Detection And Recognition In Complex Scenes

Pedestrian is the focus of video surveillance, pedestrian detection and recognition isone of computer vision’s basic tasks and key technologies. So it is widely used inintelligent video surveillance, vehicle auxiliary driving, human-computer interaction etc.This thesis tries to get insights on some key issues, such as complex scenes, cameramovement or not, pedestrian varied poses, pedestrian feature extraction and integration,pedestrian segmentation and selection of candidate targets etc, in pedestrian detection andrecognition. The main research contents are as follows:1) In the single viewpoint and fixed camera scene, This paper proposed a newalgorithm of pedestrian detection based on Kernel Density Estimation of localspatio-temporal model(LST-KDE), which overcame time-consuming of backgroundmodeling, complex scenes and poor adaptation of background updating. The LST-KDEalgorithm used K-means clustering algorithm to optimize sample set and choose keyframes in the training phase. In the updating background phase, the LST-KDE algorithmconstructed local spatio-temporal model. It not only adaptively set time window size byhistory frame information in temporal model, but also used color and texture featuresdescribed by LBP algorithm to avoid shadow problem in spatial model. The experiment inthe complex environment video demonstrated that the real-time performance and detectionaccuracy of LST-KDE outperforms recent state-of-the-art methods.2) In the single viewpoint and moved camera scene, for the problem of staticpedestrian detection failure and aerial video with small objects, varied poses and complexscene etc, this paper proposed a new algorithm of fusing multi-feature on pedestriandetection(HLS). The HLS algorithm fused pedestrian’s HOG, LBP and SIFT feature, andthen used PCA to reduce features dimension. Also, different features set different weights according to features proportion. This effectively settled the shortcoming of backgroundsubtraction and achieved good results in arial video.3) For the problem of target’s detail fuzzy, background mess etc, in aerial video whenrecognizing pedestrian, this paper proposed an pedestrian recognition based on Kalmanfiltering and saliency detection algorithm(KS-WRM). In the segmentation stage, theKS-WRM algorithm used saliency detection algorithm, which confirmed targets indetective scene. In the matching stage, the KS-WRM algorithm firstly used Kalman filteralgorithm to label candidate’s region, and then selected candidates using weighted regionmatching algorithm in the labeled region, which can avoid the problem of selectingconstant candidates under supervision environment. As a result, it not only reducedcalculation, but also improved adaptive and real-time performance.