Research On Video Compression For Visual Surveillance

By analyzing information from one or several cameras, a visual surveillance system can supervise and control fields distributed in large and complex areas. The present main research concentrate on how to realize or substitute human vision, or achieve a surveillance system with automation as much as possible, while study on compression of surveillance video is fewer relatively. In fact, the efficient compression of surveillance video with huge data volume is needed yet. Furthermore, new requirement is proposed on surveillance video coding. In addition, video compression inside a surveillance system is not a single problem; instead, it is a system problem with video pre-processing. Therefore, it is important to study compression of surveillance video. This paper includes three cues: one, motion detection in order to provide an remote alarm signal for supervisors or an indication for the start of local compression procedure; the other, motion segmentation for acquisition of surveillance objects and thereafter object-based video coding; the third, ROI video coding in view of some special surveillance cases. The main ideas of this paper are outlined as follows.1. Study of motion detection to provide an alarm signal and a start indication of video compressionBecause motion between adjacent frames can cause change in terms of DCT coefficient amplitudes and signs, the appearance of motion objects can be determined by the change of signs of DCT coefficients. It is showed that signs of DCT coefficients can be acquired from the M-JPEG code stream. In order to eliminate the impact imposed by random noise, the number of DCT coefficient sign changes caused by the noise is modeled as Gauss distribution, which can determine a threshold of motion blocks. The large number of false motion blocks brought by abrupt change of scene luminance is removed by median filtering. At last, the biggest number of motion blocks of a frame in the calibration stage is used as a threshold of motion frames. Experiments show accurate motion detection can be achieved under different quantization matrixes and/or different motion blocks thresholds.2. Study of segmentation of surveillance objects with motionThe camera noise statistical characteristic is approximated as Gauss or Laplace distribution, then an error probability is set and a threshold is determined. According to the hypothesis test theory, a constructed statistic can judge whether an object is moved or not. A statistic can be a difference d at the current location, a combination of all d inside a block or a template, then a determination on the current pixel, the central pixel, or a block is made. Experiments on different statistics and judging objects show that the test methods based on the two statistical distributions can make good results on test videos to different extent; but the test based on Laplace distribution is better in terms of overall performance. Experiments also show the proposed "determination on pixels—median filtering—macro-blocks mask" strategy is more robust.3. Study of object-based video compression for enhanced surveillance regionIn the proposed object-based coding framework for surveillance application, the background without surveillance information is coded once then sent to the receiver (with update online). Surveillance regions are enhanced at first to achieve the required visual effect. The shape information of surveillance objects is represented as macro-block coordinates. The temporal redundancy is eliminated by motion estimation within the surveillance region in the reference frame. Since the number of luminance levels is reduced due to enhancement process, the residual error of motion estimation is compressed with lossless entropy coding. Experimental results show that the proposed scheme can produce surveillance region with better contrast and significant compression efficiency at the same time.In order to adapt to the proposed object-based coding for surveillance video, an adaptive and fast motion estimation algorithm integrated with multiple advantages is proposed. These advantages include the prediction for search start points, different search patterns, half-way stop technology, and a search rule with priority. Experiments show that the proposed algorithm presents a competitive performance with low computation cost.4. Study of region of interest (ROI) based surveillance video compressionFor a kind of surveillance video produced by fixed camera and fixed surveillance region, the surveillance region is defined as ROI. The video is divided into group of frames (GOF). 3D wavelet transform is performed on each GOR The ROI of every frame is defined as a rectangular with fixed location specified by the coordinates of left top and right bottom corners. The samples between the two corners belong to ROI mask. Wavelet coefficients inside 3D ROI is scaled up. Then the GOF is coded by 3D SPIHT algorithm and thus ROI is assigned bit overhead with priority. Experiments show that surveillance regions can be reconstructed with priority and achieve better visual quality.