Research On Sliding Window-Based Pedestrian Detection

Pedestrian detection aims to automatically detect pedestrian from images or videos. It is one of the hot researches due to its wide application prospective. It is difficult due to the complexities of the pedestrian appearances and the imaging scenarios. In this thesis, we study pedestrian detection based on the sliding window techniques and propose several efficient improvements to the local binary pattern (LBP) and a new fast window fusion technique. The main contributions can be summarized as follows:(1) A multi-scale cell-structured LBP and its adjustable version:The cell-structured LBP (CLBP) contains only local information, which may be unstable. We propose multi-scale cell-structured LBP (MLBP) by combining local information and global information together. We further present adjustable multi-scale cell-structured LBP (AMLBP) by incorporating a control factor into MLBP, so that a better performed MLBP can be obtained.(2) A weighted statistical cell-structured LBP mode:To efficiently utilize local neighborhood information of CLBP, we propose a weighted statistical model of CLBP. We further propose two weighted cell-structured LBPs based on this model: variance-weighted cell-structured LBP (VCLBP) and gradient-weighted cell-structured LBP (GCLBP). Both patterns use the contrast and gradient information to compute the weight of each pixel and fuse multiple features. The new features improve detection performance significantly without increasing the dimension of feature vectors. They also show the advantage of multiple features integration.(3) A unified control-factor LBP model:We expand the idea of control factor used in AMLBP and propose a unified control-factor LBP model. The LBP features based on this model are robust to illumination variations and noise. We also introduce several new feature series (ACLBP, AVCLBP, AGCLBP and ACENTRIST) based on this model to prove the efficiency of the model. (4) A window fusion method based on fuzzy equivalence relation:Current window fusion methods are slow, therefore, a fast and efficient fuzzy equivalence relation based fusion method is proposed. This method merges multiple candidate windows as a final detection window by the fuzzy equivalence relation theories. It is fast and thus shortens the runtime of human detection.