Research And Implementation Of Pedestrian Detection System At Night Based On Template Matching

Car night vision pedestrian detection system has been developed in foreign countries formany years, this technology, however, is currently only used in a small number of high-endluxury sedan, has yet to form a mainstream application. Life and work very fast in modernsociety, driving at night is extremely common, night driver facing the light, traffic and energychallenges. Therefore, the research and technologies to improve such status is significant.Firstly, in the night vision pedestrian detection preprocessing stage, this article use TI’sFVID model to catch infrared video image; Because the collected infrared video image is inYCbCr format and pedestrians are concentrated in the center of the image area, so this articleextracted the Y component data in middle region of image as subsequent processing datasource, again using local adaptive threshold segmentation algorithm to divide it into a binaryimage, and in this process this article optimized segmentation algorithm in DM6437processor;then the binary image doing expansion and erosion operation, which filter out the noise thatmay affect the latter part of the pedestrian detection; Finally, use eight-connected componentlabeling algorithm in small image processing library VLIB provided by TI, to extract the ROIregion of interest, get the possible pedestrian ROI, provide a basis for the subsequentdetection of pedestrians.In pedestrian detection phase, through in-depth study of the infrared image matchingtheory and analysis of relevant outcomes, this article uses an infrared image scene adaptationability, a simple algorithm, easy to implement and less of the calculation in actual DM6437hardware platform, called probability template matching algorithm, using the algorithm ascore detection algorithm in car night vision pedestrian detection system. In the actual testingprocess, this paper will firstly train the pedestrian samples to the probability template libraryof different size on the PC, and near distance is96*40pix, middle distance is62*24pix,fardistance is24*10pix, each distance divided into five types of templates: go to the left of thepedestrian, left cycling pedestrians, pedestrians go right, right cycling pedestrians, forward(after) the pedestrians walk and forward (after) cycling pedestrian; then stored template datato DM6437DDR2memory in advance; Second, in the actual detection process, according tothe extracted pedestrian ROI height, pedestrian automatically divided to near distant (highthan80pix in height), medium distance (high than30pix and less than80pix in height) andclose distance (less than30pix in height) three cases, and they were detected respectively; atdifferent distances, the system automatically calls the probability template database ofdifferent size in DDR2storage for pedestrian detection calculated. In the detection false rate, this paper use haar-like feature matching algorithm to reducethe false rate. Finally, when a pedestrian is detected, the system will be marked in a differentcolor block diagram based on the pedestrian location which is in the far distance, mediumdistance or near distance, give a clear message to driver; when the pedestrian is detectedwithin the dangerous range, the system will also give warning, warning the driver to drivecarefully.Finally, the system described in this paper, use the TI’s DM6437as core processor in thehardware platform. Besides the system does the optimization in memory allocation, codewriting, DSP/BIOS cutting, key algorithm optimization, IQmath library, VLIB library and soon. After optimized system, the detection program run time from1frames/sec to15frames/sec~26frames/sec, the optimization effect is obvious and meet the real-timerequirements.