Research On Target Location Algorithm In Multi-camera Collaborative Video Environment

With the price reduction of the various monitoring equipments and related technology maturing,video surveillance systems are gradually spreading from cities to villages and towns,the application directions are extending from the the safety monitoring to all aspects of production and life:The farm monitors pigs' behaviors in the real-time,in order to detect pig birth and other special circumstances immediately;the intelligent greenhouse uses pest control monitoring system,and the intelligent home system works with the human-computer interaction,these are where the video surveillance systems to show their functions.But the existing video surveillance system uses the cameras in a divided method for target tracking and positioning,a single camera's monitoring range is limited,if the installation position is improper,it will lead to a blind corner,and it will miss the target easily duing to interferences and occlusions.This simple combination of multiple cameras are difficult to meet the needs of people on the video surveillance system,so an algorithm which can make the multiple cameras work together is very necessary,it is also one meaning of our study.Multi-camera collaborative algorithm is based on a single camera monitoring algorithm,the effect of a single camera monitoring algorithm has a direct impact on the final results.Based on the comparison of the existing single and multi-camera moving target localization algorithms,this paper proposes a multi-fusion algorithm for single-camera monitoring,meanwhile,it improves the video target localization algorithm in the multi camera environment,the main work and innovation are as follows:(1)In the single camera environment,a multi-fusion target localization algorithm is proposed,and making comparison among the commonly used hue,saturation,brightness and color distance operators.Aiming at the problem of foreground dilapidation in complex background,besides the improved operators and the fusion of multi-operators,here presents two methods by comparing the different performance of foreground under the overall / local threshold,one called torn target aggregation method which based on the centroid positions and another called gridding local background extraction method.Finally,in order to match the improved operators and take the advantages of the operator fusion method,the modified Surendra/codebook methods are proposed,and their value is proved in the simulation experiment;(2)Proposing a method to extract the vertical skeletons of moving objects,and the algorithm is used to realize the pedestrian recognition of a specific posture,which is proved to be successful.On this basis,using the human body approximate ground contact points and the vertical skeleton's highest points as homography matrixs' matching points,and realizing the human body matching in the overlapping area of the multi-cameras by the main color feature based on the non uniform quantization of hue,their cooperation can achieve the adaptive generation of the homography matrix;(3)The adaptive generation of the homography matrix based on the SIFT feature is realized by detecting the region of the moving object,which are compared with the human body approximate ground contact points' method and the vertical skeleton's highest points' method to prove the value of each other.