Research On Pedestrian Detection Based On Image Fusion

Pedestrian detection is to use of certain methods and techniques to automaticallydetect traveling people from the image or video stream, which also is an importantpart of the vehicle driver assistance system, and so on the basis of further analysisand discriminate of the behavior of pedestrians. The current pedestrian detectiontechnology is mainly detected based on the monocular vision sources, includinginfrared, visible light, radar. Due to the limitations of monocular visual access toinformation sources on the outside, the presentation of the outside world will form aone-sided. Such as thermal imaging infrared image is, you can use it during the day ornight, which can penetrate smoke, can find targets at night and in low light conditionsimaging. However, the current low precision of presence on infrared sensor and theobscure target edge is likely to cause the decline of the recognition. Visible image getgood visual effects, rich texture, edge clearer goals, but in the case of poor lightingconditions can lead to poor recognition of pedestrians. Therefore, this paper will bededicated to fuse the infrared and visible images together to take advantage of twovisual sources, to improve pedestrian target edge definition, and thus improve therecognition rate of pedestrians. The pedestrian detection system which is proposed inthis paper will include two parts: image fusion and pedestrian detection.In the part of the fusion of the infrared and visible light images, this paperproposes a method based on wavelet transform fusion, which make the low-frequencycoefficients and high frequency coefficients are separate fused. Among them, in theintegration of the low-frequency coefficients, using adaptive weighted fusion basedon the average gradient method; in the fusion of high frequency coefficients, using aratio based on signal strength and edge extraction characteristics of low-frequencycoefficients of the same layer fused to joint determination coefficient method. Theexperimental results show that this method can effectively improve the clarity of theedge of the pedestrian goals, which helps to improve the detection accuracy ofpedestrian detection stage. In pedestrian detection section, we propose the method of a fast head point’sgradient based to determine pedestrian detection window than conventional slidingwindow method which can reduce the amount of computation. The principle of thismethod is based on that pedestrian’s head is close to be round, whether wearing a hator not, the gradient of the head is similar to at about90°. The proposed method is tocalculate the whole image gradient, the point which meet at [80°,100°] gradient willbe set as a reference detection point; and then the point will be used as the centerpoint of the upper border of the detection window. Then only apply the pedestriandetection in the area of the detection window can effectively improve the detectionspeed. Experiments show that the algorithm has higher detection rate and efficiency.