| Pedestrian detection which is and important branch of image/video-based object detection hasfound widely promising applications in such areas as video surveillance, intelligent traffic,advanced man-machine interface, and so on. The main idea for pedestrian detection is basedon machine learning, in which object detection can be regarded as a kind of classificationproblem. By introducing appropriate feature descriptors, the pedestrian detection model canbe learned with large number of training samples.Due to the complex working environment and the need of high detection performance,pedestrian detection faces many challenges and has become an important research part in thefields of computer vision, artificial intelligence and pattern recognition, et al.Based on the state of the art, this thesis focused on the research work for pedestriandetection from the aspects of feature description and classifier training. The maincontributions can be summarized as the following.1. Pedestrian detection based on hierarchical HOG.In this thesis, hierarchical HOG is introduced for obejct description in a coarse-to-fineglobal-to-local manner based on multi-layer spatial strategy. Experiments of pedestriandetection based on the cascaded gentle AdaBoost show that the performance of hierarchicalHOG description outperforms that of HOG description.2. Gentle AdaBoost classifier training based on SVM feature pre-filtering.In order to efficiently improve the time performance during the training phase ofcascaded AdaBoost classifier, this thesis proposes a training strategy which combines featureset condensation based on linear SVM pre-filtering and gentle AdaBoost training based oncondensed feature set. Without weakening the general detection performance, the cascadedclassifier training performance can be efficiently improved when keeping the percentage offeature energy as85%~90%.3. Video pedestrian detection based on static cameras.By combining background substraction and the above research work, a video-basedpedestrain detection framework is finally proposed in this thesis. |
