Research On Some Key Technologies For Video Surveillance Systems

As the extension of the human visual system, video surveillance system is ofincreasing importance in many applications and has extremely significant realisticvalue today. Most recently, with the development of electronics and informationtechnology, especially computer vision and multimedia technology, communicationstechnology, numerous significant breakthroughs are achieved in the research area ofvideo surveillance. Especially, intelligent analysis and processing of video information,and video compression and transmission, which are two of the most fundamentalaspects in video surveillance, cover a number of valuable research problems which areopen and endless. In this dissertation, we investigate several key problems and proposesome new methods for video surveillance systems. Three basic problems are covered:face recognition, video enhancement and the optimization of video codec systems.Firstly, face recognition plays a key role in video surveillance. Although automaticface recognition technique has been mature, however, most existing face recognitionsystems succeeded by restricting themselves to controlled environments, especiallywhen the lighting condition changes, the recognition rate was significantly decreased.The all-weather video surveillance systems make the collection of facial image unableto avoid the influence of the light conditions, and thus it is very necessary to overcomethe bad effects of illumination change for face recognition in video surveillance. Thisdissertation firstly makes an intensive study on the illumination robust automatic facerecognition. Afterwards, based on the facial illumination model from human visualsystem, a new illumination invariant facial feature description method is proposed. Thismethod adopts contourlet transform to process human facial images in Logarithmicdomain to obtain low frequency subbands and high frequency subbands, and then keepthe low frequency subband coefficients unchanged and utilizes contourlet denoisingmodel to modify the high frequency subband coefficients, and finally applies inversecontourlet transform to reconstruct the facial images and to get the estimation of theillumination invariant components. In addition, this dissertation proposes another facial feature description method base on hybrid projection function. This method divides thefacial images into several non-overlapping blocks so that the local image distortion willless affect the recognition result, and then uses the hybrid projection function andimage entropy to extract the facial features and construct the feature vector forrecognition. Since the hybrid projection function is not sensitive to illuminationchanges and random noises, this facial feature description method is illumination robustfor face recognition. The experimental results show that both methods can effectivelyimprove the recognition rate under illumination changes.Secondly, video enhancement is one of the most important core technologies invideo surveillance. The video surveillance systems should work for24hours a day,while the poor light conditions in night bring lots of problems for the later analysis andrecognition tasks. Therefore, it is important to improve the visual perception quality ofnighttime video. In this dissertation, a new nighttime video enhancement algorithmusing wavelet transform is proposed. This method applies a new approach whichcombines the color space transform and wavelet transform to separate the illuminationcomponents from the video and makes motion detection and background estimation,then proposes a new video fusion rule for enhancement by exploiting the context from ahigh quality video captured in daytime from the same view, and finally reconstructs thecolor images and video. The experimental results demonstrate that the new method iseffective and competitive.Thirdly, video codec optimization is very important in video surveillance systems.As the new generation of video coding standard, H.264/AVC and SVC achieve highcompression efficiency. However, this performance gain comes at the cost of anincreased computational complexity. The contradictions between the massive videodata and limited storage and network resources, the demand of real-time processing andlimited computing power and the energy consumption of the embedded devices,become the bottleneck for video surveillance systems, which makes the improvementof the video coding speed very necessary. This dissertation proposes a fast inter-frameprediction algorithm applied in H.264/AVC. This algorithm predicts the candidatemodes based on moving and texture for the current macro-block to exclude the codingmode with low probability at first, then predicts the possibility and the coding order forthe candidate modes of macro-blocks, at last early terminates coding by combining the correlation and the quantization parameters. The experimental results show that theencoding speed can be improved effectively with negligible coding efficiency loss.Besides, by exploiting the characteristics of SVC, a fast inter-frame predictionalgorithm in enhancement layer of quality scalable video coding is proposed. Thisalgorithm combines the mode-distribution correlation between the base layer andenhancement layers, residuals and motion vectors, which is very suitable for theenhancement layer coding. The experimental results demonstrate that the new methodcan improve the coding speed with negligible loss and it is especially suitable for thecomplex video sequences.