Slow-Moving Objects Extraction via Adaptive Frame Differencing for Video Surveillance

Abstract

Conventional moving objects detection methods are not efficient for real-time video surveillance. Furthermore, most of them do not produce effective extraction results for certain types of moving objects: slow, fast, far, and near. This paper presents a moving objects extraction algorithm to detect the abovementioned moving objects simultaneously. This method includes moving detection by adaptive frame differencing, binarization by automatic thresholding, bounding-boxes are obtained by connected component labeling, and localization the moving objects by cascade framework. The adaptive frame differencing uses different inter-frame for frame differencing. The number of inter-frame depends on variations in the differencing image. The thresholding method uses a modified triangular algorithm and these variations to determine the threshold value and reduces most small noises. The localization cascade framework combines bounding- boxes-based morphological operations with associative rules. This framework merges broken objects and removes noise pertaining to small and spread connected components. The fps value (maximum 72) depends on the speed of the objects. The number of inter-frame is inversely proportional to the speed. The results demonstrate that our system is more efficient than traditional moving objects methods. The true and false positive rates are 97.58% and 0.438%, respectively.