Research On Pedestrian Detection Based On Normed Proposals And Key Points Matching

With the improvement of hardware performance and software quality,the computer vision has made considerable development and pedestrian detection as the most basic and most important task is always a research focus in computer vision.Pedestrian detection is well used in intelligent security,traffic monitoring,motion recognition and virtual reality,and it has important research value and broad market prospects.Over the past decade,the pedestrian detection technology has made considerable achievement in which accuracy and speed of detectors are both improved quickly.However,due to the inherent characteristics of pedestrian and the complexity of the scene,the detection effect can't meet the needs of practical applications.How to enhance the adaptability of detector in changeable scene,as well as how to improve the detector's capacity to detect pedestrian in an occlusion are still research hotspots in pedestrian detection technology.In this paper,we propose a new fast moving pedestrian detection algorithm to solve the problems in which detector may not adapt to the complex scene and may suffer from detecting pedestrian in an occlusion.The main work and innovation of this paper are described as follows.The weak adaptability of detector in change scenes is one of main restricting factor of pedestrian detection technology.We think that the movement is the commonality of pedestrian,and then propose a new flash-bit computing method to quickly obtain motion regions of interest in an image.In the method,we consider each two adjacent images in a time series as a computing node,and then filter out those obvious motion targets from background according to the differences in two adjacent images,finally divide those targets into single targets and difficult targets according to their changes in shape and displacement.When compared to the classic sliding window method,our flash-bit computing method processes images' date in a bitwise way,so it greatly improves the speed of detector and meets the real-time requirement of an application.In order to improve the ability for the detector to detect pedestrian in occlusion which names as difficult target,we constantly compute and update the BRISK key points in the single region which contains single target only.The detection algorithm focuses on the case in which a single target converts to be a difficult one due to obstructions in background or multi-overlap between two targets.When the case occurs,inspired by CMT algorithm,we match the pedestrian in occlusion by an optical flow method based on those BRISK key points saved and updated before.We name our key points matching method as KPM.When compared to CMT algorithm,our KPM which with an updating strategy of key points can simultaneously match multiple difficult targets and significantly reduce the possibility of drifting of detection windows.We use both mathematical derivation and data analysis to demonstrate a fact that normalized normed proposals have high degree of matching pedestrian.We select some normed proposals outputted by BING algorithm and then correct the shape and location of each detection window.Those detection windows after corrected improve 5%accuracy of detector successfully.Results of comparative experiences of five detectors in three motion pedestrian databases show that our algorithm achieves not only a real-time speed,but also a better accuracy that more than half of difficult targets are detected successfully.