Human posture detection for video surveillance.
Date of Issue2009
School of Computer Engineering
This thesis has made 3 major contributions. First a novel, effective and efficient framework to segment and keep track of Moving object Outline (MO) is proposed. We employ only temporal difference that takes difference of two consecutive frames and difference of the edge images followed by thresholding to extract outlines of moving regions. Compared to background subtraction, it works without referring to the background image and hence it does not need to spend valuable resources to update the background. It is suitable and convenient to be set up at non-predetermined scenes (mobile surveillance) with complex background. It is faster than optical flow method and can be applied in real time. Its effectiveness is measured by human perception when the segmentation results of 4 different types of inputs (totally 81 sequences) are displayed consecutively to 30 observers. Compared with other researches on motion segmentation, the proposed techniques are simple in concept and can work well in different kinds of environments such as dim lighting, lights changing, inside elevator and low quality images from close-circuit cameras or TV. This advocates a new direction for motion segmentation and becomes the major contribution of this work. To recognize human behaviors, pixel intensities are transformed into features to represent different actions. Hence, the analysis of feature pattern plays a vital role for human action recognition. The second contribution of this thesis is the derivation and design of useful motion features. We generate the front/rear motion history images (rMHI and fMHI) and motion vectors along the moving object boundary to measure motion and distinguish which side of outline belongs to moving region. In addition, Sedimentary Moving object Outline (SMO) for representing motion history is employed to record MO that stop moving. Hence, deposited objects can be discovered and tracked. To track movement of overlapping body parts, inner MHI (iMHI) is proposed to isolate the mixed moving parts from each other and then track the movement of each isolated part. The iMHI can be applied to generate motion feature for each tracked moving part. The third contribution proposes one approach for human posture computation, which include two sequential processes: convex shape modeling and convex shape combination. Convex shape modeling aims to generate convex shapes that can be the candidatures for human head and torso. Then the head, torso and limbs combination is located by combining two convex shapes according to the topology of the estimated body posture. The last two contributions are built on top of the first contribution of motion segmentation. Finally, the implementation has built a motion analysis system that supports video surveillance of human body motion.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision