Distributed Video Coding

Introduction

Conventional video coding methods, using predictive coding techniques to exploit the statistics of adjacent video frames, succeed in a wide range of applications. However, the predictive coding strategy makes encoders computationally intensive and thus is not suitable for emerging applications such as wireless video cameras and wireless low-power surveillance networks. These applications have limited processing power and battery lives.A new coding paradigm, Distributed Video Coding (DVC) provides an inverse complexity balance between the encoder and decoder. This novel feature is ideal for solving the above problems. A DVC codec typically divides the video sequences into key frames and Wyner-Ziv (WZ) frames. They can be separately encoded without any reference to each other, but still achieve similar coding efficiency in terms of rate distortion (RD) performance as the conventional coding approaches. In addition, video transmission over wireless links is unreliable and can be characterized by bursty and high channel error probability. Since channel coding is adopted in DVC, this brings another appealing property that it is resilient against transmission errors. 

System Architecture

• The input video is divided into key frames and WZ frames. Key frames are inserted periodically determined by group of pictures (GOP) size.
• The WZ frames are further divided into 4-by-4 blocks and in each block, discrete cosine transform (DCT) and a uniform quantization with deadzone are performed. 
• The quantized DCT coefficients are grouped into frequency bands and converted into bit-planes.
• Each bit-plane is separately encoded using low-density-parity-check (LDPC) codes and stored in a buffer for decoder requests. An 8-bit cyclic redundancy check (CRC) code is also generated for each bit-plane to confirm decoding is success. 
• Key frames are encoded by an efficient conventional coder such as H.264/AVC Intra.
• At the decoder side, decoded key frames and WZ frames are interpolated to generate SI.
• The correlation noise between SI and WZ frames are assumed to be Laplacian distributed and modelled by the virtual channel model.
• Distortion of AC coefficients is estimated by residual statistic information of the decoded key frames and sent back to the encoder to aid WZ frame quality control. These results are further utilized in the virtual channel modelling process.
• The soft input to the LDPC decoder in terms of conditional bit probability is calculated, using the statistical information provided by the virtual channel model. 
• An iterative decoding process is performed until the syndrome check and CRC check are both successful. More parity bits can be requested if the above stopping criterion is not met.
• Finally, all the decoded quantized symbols are optimally reconstructed and inverse transformed.

DSP Implementation of On-Board Distributed Video Coding

          The first implementation of a distributed video encoder on a Texas Instruments TMS320DM6437 digital signal processor (DSP) is proposed.  The encoder consists of a Wyner-Ziv (WZ) encoder and a conventional intra-frame encoder. The WZ encoder is efficiently implemented, using rate adaptive low-density-parity-check accumulative (LDPCA) codes, exploiting the hardware features and optimization techniques to improve the overall performance. Implementation results show that the WZ encoder is able to encode at 134M instruction cycles per QCIF frame on a TMS320DM6437 DSP running at 700MHz. This results in encoder speed 29 times faster than non-optimized encoder implementation.

Exploration and Exploitation of Reference Frames

          This work investigates the impact of reference frames in DVC on the Rate-Distortion (RD) performance and reveals that reference frames have the potential to be better side information (SI) than the extensively used interpolated frames. Based on this investigation, we also propose a motion range prediction (MRP) method to exploit reference frames and precisely guide the statistical motion learning process. Extensive simulation results show that choosing reference frames as SI performs competitively, and sometimes even better than interpolated frames. Furthermore, the proposed MRP method is shown to significantly reduce the decoding complexity without degrading any RD performance.

Consistent Quality Control for Wireless Video Surveillance

          DVC is well known for low complexity encoding which provides coding solutions for a wide range of applications, in particular wireless video surveillance. We address the problem of distortion variation introduced by typical rate control algorithms, especially in a various bit rate environment. A distortion quantization model derived from a MPEG-2 distortion estimation model is proposed to achieve smooth picture quality across video frames. Simulation results show that the proposed quality control algorithm is capable to meet user defined target distortion and maintain a rather small variation for sequence with slow motion and performs similar to fixed quantization for fast motion sequence at the cost of some RD performance.

Full Low Complexity Implementation of CRG-DVC Codec

          A full low complexity implementation of the CRG-DVC codec, which consists of a PC based encoder and a parallel implementation of the decoder based on a HPC (high performance cluster) is proposed. 

Figure 2 - RD Curves for DSP-PC based DVC codec implementation for different sequences.