Pedestrian Detection Using Combinations of Multiple Features and Object Tracking Method

Pedestrian detection has become a very important topic in computer vision, due to the higher attention paid to driving security. This thesis aims at building a platform for real-time pedestrian detection, using Central Processing Unit (CPU) of personal computers to process the video stream data for pedestrian detection.
This thesis first reviews some research works of pedestrian detection done by Digital Camera and Computer Vision Laboratory, Department of Computer Science and Information Engineering, National Taiwan University and also mainstream detection method of finding pedestrians in an image. We significantly reduce the miss detection rate by some combination of features of Histogram of Oriented Gradient (HOG), Local Binary Patterns (LBP) and Color histogram of Self-Similarity of low-level features (CSS).
Not only to reduce the miss detection rate, our program also increased the detecting speed to achieve the real-time performance. This thesis analyzes the relationship between search regions and the height of pedestrians by assuming driving recorders are placed at a specific angle towards the ground and the camera to be a pin-hole model. Besides pedestrian detection, we also apply the Lucas-Kanade method for computing the optical flow to track the pedestrians who have already been detected between two consecutive images and significantly reduced the time consumption. Meanwhile, for HOG features, we also apply a method to infer features under different scales to save the computation time of integral feature images. By using the methods above, we build a platform for real-time pedestrian detection.