Research And Application Of Human Tracking Based On Multi-Cameras

Intelligent video surveillance technology was a topic of active research and was draw wide attention of the public in recent years.Multi-cameras human tracking is important component in intelligent video surveillance.There are two key problems in multi-camera tracking: Tracking algorithms that works well has slower processed speed.Appeared a specified pedestrian in some camera,can locate the same person in other cameras automatically.Aimed at the two problems that researched in this paper.The contents of this paper are:(1)In order to solve a problem that can't track real time in high resolution,proposed a method base on circular queue and multi-thread.First,create two circular queues as buffer,used to store frame of have not processed and a frame of scaled.Next,one thread used to read frame to buffer and scaling frame,the other thread used to process frame of scaled using tracking algorithm.At last,Get location information from tracking algorithm and mapped to frame of not processed to show.In order to avoid some situation that overlaps the data in circular queue due to thread is so fast,can increase the capacity of circular queue.Due to different operations,the speeds of process are different,need to do something according to situations.In order to protect the consistent of show-frame and process-frame,same threads operate on different circular queue are should be FIFO and atomic.The experiment shows,in high resolution,can track object real time.(2)In order to judge human images from other cameras is consistent with current camera human image,proposed a match algorithm base on color histogram.First,Count H and S component in HSV color space of all images to remove the effect of brightness in image.Next,Normalized the 2D color histogram after extracted,avoid the number of resolution are not same in different images.At last,compute the match values in different images.This paper has tested in VIPeR dataset,and gets good result.(3)Due to global feature is easy to interfere,proposed a human match algorithm based on SIFT in this paper.First,due to the result of human body detect contain background component to interfere.Threshold the human body images,and then detects contours.The bounding box of outmost contour is body region,Next,detected keypoints and generates descriptor in the region.And get all the information of matched,and then give a weight to several keypoints that nearest.At last,calculate the sum of squares.The experiment shows,the method is valid.(4)Experiments are made by OpenCV3.1 and CUDA7.5.Due to CUDA module,the program is 64-bit.Shoot some videos to simulate the situation of multi-cameras and track human.The experiment result shows can track human body validly and remarkable effective.