| The service of video surveillance has a very long history, which is widely used for security to keep the society safe. Currently, with the development of computer technology and image processing , video surveillance is widely used in education, government, entertainment and sports domain.Video coding is one of the indispensable component and foundation for a digital video surveillance system. It can provide benefits for image transport and video storage. The technologies used in the current video coding standards, such as quantization, transform and entropy coding, are mainly designed aiming at video taken with good light conditions, which has little noises in the images. Thus, the output bitrate of coding night video using these standards increase dramatically comparing to that of coding normal video due to various noises existed in the night video. Most main research efforts now deal with such problems by reducing noises using image enhancement techniques. However, it can bring more distortion to the recovered night vision images. Some interested detail information in the image may be removed away with noise, which is not acceptable for security video surveillance.Night vision video is a type of low light level (LLL) videos. One common characteristic of most of these videos is that there exists various noises inside the video, which limit the lowest illumination of the system and yield the image appearing with random glitter. Yuval S. Boger has testified that random glitter in LLL videos generally lasts for the time of one frame at least.In this paper, we propose a new video coding method for night vision based on analysis of image characteristic, which can keep detail in the image and reduce the coding bitrate at the same time.An experimental analysis of a typical night video, which is night_static sequence, and a surveillance video at a cross street, which is cross_street sequence, is provided in order to show some detail features of night video. Both the two videos take the YUV format 4:2:0, which is the most widely used format in practical video coding system. We compute the block correlation of every two adjacent Y frames, two adjacent U frames and two adjacent V frames in the whole video sequence respectively. PSNR (Peak Signal to Noise Ratio) is used to evaluate the correlation. For every two adjacent Y frames, we calculate the PSNR between each pair of 16×16 blocks with the same spatial location. We assume the number of block pairs with PSNR greater than a certain threshold T for the whole sequence is n, the total number of these 16×16 block pairs is N, then (n/N )×100% gives the percentage of 16×16 block pairs with PSNR greater than T. For U and V frames, the block size is 8×8.In general, the night video is full of random noises and most blocks in the video appear darker color due to weak illumination. According to the statistical data, we can see that noises dominate in the U frame comparing to Y and V frames. For the typical night video sequence night_static, this feature is very clear, while for sequence cross_street, which has better illumination, the feature is not very clear due to less random noises existed in the video. As far as Y frames in the night video are concerned, there are two categories of 16×16 blocks. The first category is with the dispersed histogram. The second category is with the centralized histogram.Obviously, these pixel values are far away from most pixel values in the macroblock. They may be either noise or the detail area of the image. Anyway, it is just these pixels which cause the increased coding bitrate of the night video, because more bits are needed to encode the 16×16 block with these isolated pixel values.Based on the characteristic analysis of night video, the processing of Y, U and V frames are different in our proposed method. Because human eyes are more sensitive to illumination than to color information, so considering Y frame of the night video, the sectional function mapping is performed before the frame is encoded. In the decoder, the Y frame is mapped back by using inverse sectional function after normal video decoding.In order to preserve the detail when reducing random noises from U frame, we do not run noise reducing algorithm for all blocks of U frame. Since Y, U and V frames have the same texture for clear video, that is, the flat regions and edge or detail regions of one frame corresponds to those of another frame, we use the characteristic of the block of V frame to detect if the corresponding block of U frame contains detail information. When PSNR between 16×16 blocks of adjacent Y frames is less than a certain threshold, the corresponding 8×8 blocks between adjacent V frames presents some characteristic. In the implementation, we use the average of two adjacent V frames to detect the characteristic, because the average V frame may have less random noises, which is better to detect if the corresponding 8×8 block of U frame contains detail information. The PSNR of corresponding 16×16 blocks between two adjacent Y frames is less than 29.0 dB, which implies the corresponding 16×16 blocks of Y frame mostly are with edge or detail. We design a very simple algorithm to reduce noises from blocks of U frame in order to be applied in the real time applications such as video surveillance.All experiments are carried out under VC++6.0 platform with C language. To evaluate the performance of the proposed method, we use H.264/AVC JM10.2 as the normal encoder and decoder. We compared the coding efficiency of H.264/AVC with the proposed method. The experimental results show that the proposed method can decrease the encoding bitrate effectively under the same reconstructed video quality. The results of night_static sequence show that the proposed method can reduce bitrate over 30% compared to H.264/AVC under the same PSNR_Y. |
PDF Full Text Request