The Research Of Pederstrian Detection And Tracking Technology

With the rapid development of computer technology and artificial intelligence technology, researches about computer vision and pattern recognition become increasingly popular. Pedestrian detection and tracking technology, as one of the most popular direction in it, is of research values and practical significance in fields such as intelligent transportation system, unmanned vehicles and public safety. A brief research content of the thesis will be summarized as follow.At first, the thesis will focus on the extraction of pedestrians' region-of-interest.Pedestrian detection conducted at early times was based on pictures with low pixels. Pictures in today, however, are usually high-definition and contain more information, which would greatly increase operand and result in failure to achieve instantaneous effect. The thesis also put forward detecting method based on pedestrians' region- of-interest to solve the problem that detecting of pedestrians is too slow. The thesis will research extraction of pedestrians'region-of-interest under two different cases, namely, single-frame images and moving video sequence images. Adaboost algorithm and cascade classifier will be applied to single-frame images and pedestrians' region could be extracted quickly through reasonably resetting the lower limit of detection rate and upper limit of false drop rate in each class. As far as moving video sequence images are concerned, the thesis will improve traditional symmetrical differencing to eliminate the generating of empty holes, to precisely get the profiles of moving pedestrians and to further realize the fast extraction of pedestrians' region. The result of this experiment indicates that this kind of algorithm can quickly extract pedestrians'region-of-interest, which greatly shrinks the range of subsequent pedestrians' detection and make it possible to realize instantaneous detection.And then the thesis will study about the issue of pedestrians' precise detection on the basis of the extraction of pedestrians' region-of-interest. According to relevant researches,HOG was proved to be the best method to describe pedestrians' features among all the methods used in pedestrians' detection. Therefore, the thesis will adopt an algorithm combining HOG with SVM to conduct precise detection to pedestrians. We can see that the algorithm is of high detection rate while being likely to cause slow detection and small proportion of false detection at the same time through analyzing principles of HOG features'extraction and SVM classifier's classification. Then pedestrians in the region-of-interest would be detected on the basis of previous extraction of pedestrians' region-of-interest. The result of the experiment reveals that while keeping high detecting rate at the same time, the algorithm adopted in the thesis could reduce false detection rate and raise the speed of the detection, reducing detecting time for per frame from 1.8 seconds to 0.19 seconds.At last, the tracking of pedestrians would be studied on the basis of pedestrians' detection.After introduction to relevant theories about continuous energy minimization, the study would focus on the constructed model of energy function and build a unified energy function including pedestrians' observation models, dynamic models, repulsion models, track maintenance models and track correction models. Correlation tracking would be combinedwith policy-making tracking to solve the problem of occlusion in the movement of object.Gradient descent would be applied in the end to solve the constructed energy function and hop would also be brought in to seek optimal solution. The result of the experiment would indicates that the method adopted by the thesis could achieve precise and stable continuous tracking under the circumstances that pedestrians are partially occluded and the direction of motion is uncertain.