earticle

영어

Frame difference is a quick dissimilarity based segmentation approach for object detection, unfortunately, it gets trapped in over-segmented when the pixels of interest over time overlap each other. This paper presents a rather fast visual object detection approach capable of approximating the location of moving object under heavy background noise or big overlap caused by negative similarity. Specifically, frame forward-backward difference concept is proposed to extract object features in current frame through fusion of pixel-based current-previous and current-following frame difference. Based on this, we formulate object localization applying the statistics of horizontal-vertical projection of the fused difference. Therefore, our object detection can be regarded as a direct thresholding process which guarantees high efficiency while holds good accuracy performance. We evaluate our method on Weizmann human action dataset and some traffic videos for both single and multiple objects detection which demonstrates its applicability and prospect.