A novel hybrid human detection system

This paper describes a novel system for hybrid human detection. The detection is based on combined human models, such as head-shoulder model, face model, skin color model, hair model, and histogram of oriented gradients (HOG) model. Adaboost technique is adapted to build a strong classifier to integrate multiple human models as listed above. Our system consists of three parts. First, moving regions are detected and segmented from the input video sequence using improved mixture of Gaussian models (GMM) [1]. We improve the conventional GMM method by eliminating smearing effect, caused by slowly moving objects, and accelerating the process of GMM method. Second, shadows are removed by adopting shadow pattern in HSV color space. Third, our hybrid human models are used to detect humans in the pre-segmented moving regions. A real-time system has been implemented to segment and classify objects. Examples of object classification (humans, running dogs, vehicles) are provided in this paper. Experimental results show that our hybrid models can effectively detect humans in video sequences from fixed monocular color or grayscale camera, and robust to observation noise, lighting changing, etc.