A Real-time Multi-target Detection Algorithm Based On SVM

Moving target detection and tracking play a very important role in intelligent videosurveillance, and the corresponding technology is widely used in the field of intelligentsecurity, crowd counting and so on. Wherein the detection of moving targets is the most basicand most critical technologies over the entire video process and the effect of its detectiondirectly impact on the success of the follow-up information processing.We take advantage of the SVM detection algorithm based on histogram of orientedgradients (HOG) since the goal of this paper is to detect pedestrian objects. However, theconventional SVM detection algorithm needs large computation, which is difficult to meet thereal-time requirement in practical application. In this regard, this paper combines with targetdetection algorithm based on motion information to extract the area of moving target, therebyreducing the searching process of SVM algorithm.This paper presented the following improved algorithm in two modules: the areaextraction of moving target and the SVM detection algorithm, for the speed and accuracy oftarget detection.(1) Use an average background model with a variable weighted value. Considering the speedof the background and the effect of background extraction, we choose an averagebackground model with a variable weighted value which requires minimal computation.The weights are updated for each pixel according to the type of the model in which thepixel region dependent, therefore the extracted background is more real;(2) Improved connected component algorithm. Use region growing method based on lengthcoding to solve the connected component problem of the binary image, compared withthe traditional pixel-based region growing method, reducing the amount of redundantneighborhood judgments, speeding up the speed of growth of connected component,thereby reducing algorithm running time;(3) HOG feature extraction based on entropy. When extracting HOG features in local area,making full use of the distribution of the gradient direction, calculate the correspondingentropy value, and that value as the weight multiplied into the corresponding HOGfeature vectors. Experiments show that the HOG feature vector combined with entropycan enhance the discrimination ability of SVM classifier;(4) Target detection of SVM in discontinuous frame. Since the target shape is generally notmuch change in several adjacent frames, so no need to carry out SVM target detection ineach frame. The strategy of this article is taken once SVM detection every five frame, which can ensure the accuracy of SVM detection, but also accelerated the overall speedof SVM detection.