Research And Implementation Of HOG Based Human Detection In Image

Human detection is the technology on researching how to make computer think human-like and mark the human object region in image and video. It can be widely used in national defense, public security, electronic games etc.Classification based human detection includes several steps. First, scan the input image by detection windows of varient sizes. Then, classify detection windows. Last, generally in detection, multi-results might be generated on one single object. So must fuse them to get final detection results. The performance of classifier has close relationship with feature it chooses and classifier training. There are abundant features can be chose: color, intensity, and some local image features. Owing to the colors and intensities vary a lot; the most suitable feature could be Histogram of Oriented Gradient (HOG) which describes the edges of human object. Its extraction processes include: color generalization, gradient computation, orientation and spacial binning and high-quality normalization of over-lapping blocks. Another element contributing to the performance of classifier is classification methods. Support Vector Machine (SVM), a simple binary classifier, is widely used. These two elements make the HOG and SVM based human detection algorithm. When there are many regions needed to be computed repeatly in feature extraction, integration method can be used. It could shorten computing time and circumvent the disadvantage of lacking global information. Meanwhile, cascade performs well in classification. It consists of several levels of SVMs which makes feature match faster but enlarges classifier training time and computation. There are many result fusion methods. The algorithm of Non Maximum Suppression can fix them by mapping multi-results into 3D space, evaluating their tensity and then getting the best one.The HOG and SVM based human detection method could detect human of variant poses in complex backgrounds. While the Integrated HOG and Cascade based detection method accelerates feature extraction and classification, while maintaining the accuracy.