Based On The Characteristics Of The Study Of The Human Body Detection

Pedestrian detection is a very important branch of object detection. It has beenwidely used in smart surveillance, driver assistant systems, human-computer interactionand so on. Due to such tremendous research value and application prospect, pedestriandetection has become a popular research area in the field of computer vision. However,it’s still a challenging task for the diversification of posture and appearance, theinterference of background, lighting condition and occlusions.The prevalent approach istransforming the human detection problem into a binary classification problem whichincluding the procedure of training and detecting. The first step in training process isextracting the effective feature, and then using the statistical learning method toestablish the human body classifier. In detection part, the classifiers are used to scan theimage to find the human body. This method is adopted in our paper.Research status and difficulties in pedestrian detection are summarized in thispaper. Then three human feature sets-Harr-like feature, HOG feature and LBP featureare introduced in detail and the performances between HOG feature and LBP featureare compared. Afterwards, we combine both to propose a new HOG-LBP featurewhich is extracted based on blocks in image. Detector trained by the combinationfeature is compared to the one trained by a single feature using support vector machine.Furthermore, we discuss the performances of detectors trained by different division ofblocks to find the best method.At last, we use the best classifier to detect pedestrians. In the procedure ofdetection, multi-scale “image pyramid” and sliding windows are adopted. Consideringthat one object may correspond to multi-results, fusion method is taken to get then finalresult.Experiments show that our HOG-LBP feature can adapt the diversification ofposture, appearance and complex background and is very efficient in representingpedestrian. Our detector is more discriminative and robust than the state-of-the-artalgorithms.