A central theme in computer vision is to describe an image in terms of the meaningful objects that comprise it (such as persons, airplanes, etc). It is the basic for computers to automatically achieve the visual perception of the real world. Extracting suitable features for image data representation is one key step of object recognition. This thesis mainly studies object detection (especially face) in static image and object classification in video, especially studies extract vision features of objects with good ability of discriminative and low compute complexity. The contributions of this work are as follows: (1) In the domain of face detection, Haar-features, which are widely used, seems too simple, this results in high computer complexity both in training procedure and detection procedure. In this thesis, we use MB-LBP feature to replace Haar-feature. In our experiments, MB-LBP features show more distinctive performance. MB-LBP is extended from the original LBP feature. Compared with the original LBP, MB-LBP can capture image structures with different scales and aspect ratios. Aiming at dealing with the non-metric feature value of MB-LBP features, multi-branch regression tree is developed to construct the weak classifiers when applying AdaBoost algorithm. (2) This thesis describes an appearance-based method based on MB-LBP features to classify objects in video. This method achieved real-time and robust objects classification performance in diverse camera viewing angles. Besides classifying objects to human or vehicles, we also studies to classify the objects into car, van, truck, person, bike and group of people. The ECOC-based method is introduced to solve this multi-class classification problem. (3) This thesis studies use some shape-based and motion-based features to improve the moving object classification performance of MB-LBP feature based classifier.
修改评论