| Clutter background and partial occlusions are two challenging tasks of pedestrian detection in real world environments. It is important to solve these problems so that pedestrian detection can be put into practical applications. The technologies used in pedestrian detection can also be transferred to other object detection fields, such as face detection, and vehicle detection, and promote the development of computer vision and pattern recognition. So we focus on how to attack the problems of clutter background and partial occlusions in this thesis.As to the problem of clutter background, we make a comprehensive study of the Gaussian Mixture Mode (GMM), and then improve it. The experimental results show that our improved method is effective. A shadow detection scheme based on chrominance distortion is also introduced. We also use seeded region grow method to solve a problem of GMM—when the direction of moving objects is vertical to the imaging plane of camera, it can only detect the parts of the objects.With respect to the problem of partial occlusions, we describe a novel method based on Implicit Shape Mode (ISM). Following a common consensus in the field of object categorization, we do not assume that a figure-ground segmentation is available prior to recognition. The combination of recognition and segmentation into one process is made possible by our use of ISM, which integrates both into a common probabilistic framework. In addition to the recognition and segmentation result, it also generates a per-pixel confidence measure specifying the area that supports a hypothesis and how much it can be trusted. With this confidence, we derive a extension of the approach to handle multiple objects in a scene and resolve ambiguities between overlapping hypotheses with a novel Minimum Description Length (MDL) based criterion.Finally, we proposed a shape-based classification method for pedestrian. It unwraps the contour of a pedestrian into a one dimensional distance, and then uses Expectation Maximum (EM) to model it. The experimental result shows that there are only 27 false-classified samples in 4007. |
