Smart Surveillance System for Cattle Detection

Abstract

Video object segmentation is an important part of real time surveillance system. For any video segmentation algorithm to be suitable in real time, must require less computational load. The dissertation work presented here is divided into two main parts: (1)moving object detection, (2)identi cation of cattle from detected moving objects. To detect a mov- ing object the rst step required is segmentation. Three di erent segmentation methods namely K-means clustering algorithm, region growing algorithm and background subtrac- tion method, have been applied on di erent videos and the results were analyzed. It could be found experimentally that the computation time required for K-means algorithm is very large. Therefore it cannot produce results in real time. Region growing algorithm requires the seed points which indicate the moving object. Hence region growing algorithm is also not applicable in real time. Background subtraction method is fast and produces good results. But it requires a background image from which subsequent frames will be sub- tracted to obtain the moving object. In order to adapt the background image with the changing video scene, a new algorithm for generating an optimal background image has been proposed. Adaptive background generation is done by evaluating the probability of occurrence of the intensity value at each pixel coordinate. Algorithm has been implemented on videos with various changes in the scene in which the results are quite encouraging. Re- sults obtained with the proposed algorithm are compared with the traditional background subtraction method. For identi cation of the cattle feature based method has been used. The algorithm is applied on binary image which is obtained after performing background subtraction. By considering the changes in intensity of the neighboring pixels, a count is maintained which is incremented every time the intensity of the pixel changes from 0 to 255. Keeping a condi- tion on the number of white count obtained can identify whether the moving object under consideration is identi ed as a cattle or not. The algorithm has been applied on di fferent videos and satisfactory results were obtained.