In this paper

we present an approach for detecting pedestrians from moving backgrounds which is based on compound features combined with motion and static features. It is difficult to discriminate human bodies from a moving background. We improve Nagel's second-order gradient optical flow algorithm and enrich the inner repeatability of MBH (motion boundary histograms) and IMH (internal motion histograms) motion feature based on the flow. We train a linear SVM (support vector machine) classifier using features made from a pedestrian sample. A fixed window sliding over image and classify results are optimized using the mean shift algorithm. The accuracy is 98% on test of 1 093 group images

which is better than the results using other methods.