Research On Human Head-Shoulder Detection Based On Omega-Like Profile Feature

Human detection has become an important research topic in the field of intelligent video analysis. It is widely used in intelligent monitoring system, human-computer interaction, virtual reality, etc. However, actually due to the comfortless position and fixed angle of the camera, it is difficult to obtain an entire and upright body profile, especially when partial occluded, or a wide range of variations in pose, appearance, clothing, illumination and background. Therefore, the robust feature of head-shoulder profile which is an omega-like shape, became a breakthrough of body detection, and was discussed in depth throughout this thesis.The pedestrian motion detection in complex environment, as well as the building of head classifiers based on the head-shoulder feature were focused in this thesis. The framework of this article and the innovations are as follows:Firstly, the original collected images were preprocessed by employing Fast Median Filtering to reduce noise, and grayscale stretch to improve the contrast, to be conducive to post-processing.Secondly, with detecting moving object, the foreground containing pedestrians was extracted out to shorten scanning time. A codebook background model was constructed for motion detection. Considering the pixel value directly related to the brightness criterion, the RGB color space was converted into YUV to improve the testing performances.Thirdly, two cascaded classifiers were designed, one based on Haar-like feature for preliminary filtration, the other based on HOG feature for precision filtration. It is worth mentioning that, because of an obvious symmetry of head-shoulder part, a new feature extraction method, called Joint HOG was proposed, and blocks in multiple scales adopted as well to add basic HOG features, so that it overcomes the weakness of a single feature and time consumption. The HOG-based classifier cascade was trained by using linear-SVM vs. Discrete AdaBoost (AB) algorithms, where linear-SVM for the weak learning rule to create basic Joint HOG classifiers, and AB can select weak classifiers those of better distinguish abilities, to construct a final strong SVMs-cascaded classifier. It has two advantages: good generalization performance to estimate a high-dimensional HOG vector, and AB boosted the classifiers by choosing stronger ones together. Finally, by scaling the frame image passed each dense scanning, the two cascades of Haar-like and Joint HOG, were successively used to evaluate through every sliding detecting window. According to the classifier threshold preset, to merge the neighboring candidates, and head-shoulder objects were marked out overall the image.