Abstract
Researchers at Tsinghua University have developed a novel video encoding approach specifically for "smart eyes"—power-constrained devices such as wireless surveillance cameras and mobile sensors. By leveraging Underdetermined Blind Signal Separation (UBSS), the proposed framework shifts the heavy computational burden of video compression from the encoder to the decoder. This architecture achieves significant energy savings and faster encoding times by treating consecutive video frames as mixed signals that are separated at the receiver. Experimental results demonstrate a 50% reduction in energy consumption compared to H.264 and significant speed improvements over existing low-complexity codecs.
Introduction
Traditional video standards like H.264 are "encoder-heavy," relying on complex processes such as motion estimation and discrete cosine transforms (DCT). While effective for high compression, these processes consume substantial power, making them unsuitable for battery-powered smart cameras. Alternative frameworks like Distributed Video Coding (DVC) and Compressive Sensing (CS) often face a trade-off between quality and complexity or suffer from slow decoding. This paper proposes a middle ground: an encoding process primarily consisting of simple matrix multiplication, optimized for hardware efficiency.
The UBSS Framework
The core methodology is inspired by the "cocktail party effect," where individual signals are extracted from a mixed source. In this framework, multiple video frames are mixed into a single "mixed frame" at the encoder using a random Gaussian mixing matrix. Because there are more source frames than mixed observations ($n > m$), the problem is mathematically "underdetermined."
The video sequence is divided into Key Frames, encoded via standard H.264 Intra, and UBSS Frames, which are the mixed sequences. At the decoder, the TVAL3 algorithm (Total Variation Augmented Lagrangian) is used to separate the mixed signals and reconstruct the original frames by exploiting temporal correlations.
Key Optimizations
To improve performance and feasibility, the researchers introduced three critical optimizations:
- Residual Mixing: Instead of mixing raw frames, the encoder subtracts the preceding key frame and mixes the resulting residuals. This increases signal "sparsity," which is vital for effective reconstruction.
- Streamed-Way Mixing: To minimize memory usage, frames are processed column-by-column as they are sampled, rather than waiting for an entire sequence to arrive.
- Block-Level Mixing: Processing frames in 32x32 blocks was identified as the optimal balance between reconstruction quality (PSNR) and decoding speed.
As shown in Figure 2, the use of residual mixing provides a massive improvement—approximately 10dB—in recovered image quality (PSNR) compared to mixing raw frames. This "sparsity" is the key to the system's high-quality reconstruction.
Results and Discussion
Evaluated using the standard "Foreman" QCIF video sequence, the UBSS+H.264 framework demonstrated superior performance in several areas:
- Speed: The encoder is 60% faster than H.264 "No Motion" and 30ms faster per frame than the DISCOVER codec.
- Energy: Simulation results indicate a 50% energy saving compared to standard H.264.
- Hardware Footprint: The UBSS encoding unit is extremely lightweight, requiring only 980 equivalent gates, making it ideal for FPGA or ASIC implementation.
- Quality: At high compression ratios (above 30), the PSNR of the UBSS method exceeds H.264 Intra and is comparable to complex DVC methods.
While the TVAL3 algorithm introduces some inherent errors at low compression ratios and decoding latency remains a factor for real-time applications, the approach offers a promising path for ultra-low-power video sensing.
Conclusion
By inverting the traditional complexity of video codecs, the UBSS-based encoder provides a viable solution for energy-constrained smart eyes. Future work will focus on optimizing the separation algorithms to reduce decoding latency and improving robustness for high-motion video sequences where residuals may be less sparse.
Summary updated