The first research in the laboratory I participated in was to realize an algorithm of Pedestrian detection. In feature extraction, we calculated Histogram of Oriented Gradient (HOG), which is widely used for its effective adaptation to changes of object pose and illumination. Then we proposed a new ensemble classifier based on L2-norm minimization principle to detect human body from static images. The proposed classifier can achieve margin maximization, but with less complexity compared to the well-known Support Vector Machine (SVM) classifier. Experimental results show that, the classifier gets a better performance in terms of detection rate, while 6-8 times faster than SVM. Now I am trying to setting up a similar detection method in another platform, car detection. My current work is optimizing feature extraction from detected images. While we in the former work set a series of detecting window in different size to contain target in varied volumes, I am trying to change the size of detected image in pyramid instead of detecting window. Thus I expect to improve the speed of detection. In further working and studying, combined with my project, I found a lot of problem confusing me. In the extracting feature from the target image, when the pedestrian is smaller(not much) in size than the training simple, the result will not be good enough. I hope to solve the problem by zooming the image with method of interpolation instead. And if more than 1/3 of body is blocked out, it won’t be detected. As for cars detection, though the complexity of the feature seems lower, it turns out that the similar method is less robust to the varying viewing angle and blockage. I am trying to add the training sample to adjust these changes. These are the problems I am solving.