Research On Video-based Pedestrian Counting Algorithm

The development of computer vision has achieved great progress in recent years,which makes it possible that computers solve various problems instead of human, whichwas finding its chief development in intelligent monitoring systems. With application oftheories in image processing and pattern recognition, these systems can recognize objectsin the scene automatically, and then make a response according to their requirements.Compared with the traditional way of monitoring the scene which is accomplished byhuman resource, intelligent monitoring systems can run for a long time without a interval,saving lots of human and financial resources and avoiding the inaccuracy caused byhuman fatigue. Under this background, pedestrian counting technology began to grow. Itcan be widely used in shopping malls, subway stations and so on. Given its huge marketpotential, this paper carried out a series of research on this technology. By using thealgorithms in image processing and computer vision, the new pedestrian countingalgorithm is realized by multiple information fusion.The method of video-based pedestrian statistics is related to the target detectiontechnique and object tracking technique in pattern recognition. In the aspect of targetdetection, histograms of the oriented gradient with support vector machine which arebased on the statistical learning theory are widely used, mostly extracting features inthe whole image and needing a large mount of calculation. Considering that the targets inthe actual scene are moving, the edge of them was obtained by ways of sobel andinter-frame difference. And then features were extracted at the points around the edge,which reduced the mount of computation and removed some false targets detected in thebackground. Detection is the base of the whole algorithm, so a lower threshold was set tomake sure that targets could be recognized as many as possible, which led to a lot of falsetargets. By using the spatial information constraints, some were removed successfully. Inthe aspect of object tracking, the paper proposed an accurate target tracking method whichwas accomplished in the way of combining detection and tracking with temporal and colorconstraints, which ensured the algorithm has a high detection rate and a low false alarmrate. Compared to other algorithms, the method proposed in this paper has thecharacteristics of high statistical accuracy, good robustness, processing images rapidly.