Abstract
Stereo vision is a critical methodology for obtaining depth information from a scene using a stereo image pair. This paper introduces a Discrete Wavelet Transform (DWT)-based approach for disparity estimation that significantly improves performance in terms of speed and computational complexity. By applying DWT to input images, the number of samples to be processed is reduced by 50%, leading to a 52% boost in frame rates and a 60% reduction in hardware resource usage on FPGAs. The proposed architecture was prototyped on a Xilinx Virtex-7 FPGA and evaluated against standard Middlebury Benchmark image pairs, demonstrating high efficiency with minimal loss in accuracy.
Introduction
Stereo matching—calculating depth by comparing images from two slightly offset cameras—is essential for autonomous drones, robots, and self-driving vehicles. However, engineers often face a "trilemma" between high accuracy, high speed, and low power consumption. While global matching algorithms are accurate, they are computationally heavy. Local matching algorithms, particularly those using Guided Image Filtering (GIF), offer a better balance but often struggle with real-time high-resolution video. This research addresses these challenges by reducing the computational complexity of GIF-based disparity estimation for efficient hardware implementation.
Methodology
The core innovation involves integrating a Discrete Wavelet Transform (DWT) at the start of the pipeline and an Inverse DWT (IDWT) at the end to manage data volume efficiently.
- Wavelet Decomposition: A 1D Haar wavelet transform decomposes input images into "approximate" and "detailed" coefficients. This process retains essential features while halving the sample count.
- Cost Volume Construction (CVC): The system calculates the "cost" (difference) between left and right pixels using Sum of Absolute Differences (SAD) and Gradients.
- Cost Volume Filtering (CVF): The Guided Image Filter (GIF) smooths the cost data while preserving sharp edges. Because the input data was halved by the DWT, this resource-intensive stage runs significantly faster.
- Disparity Selection: A "Winner-Take-All" (WTA) approach selects the most likely depth for each pixel.
- IDWT Up-sampling: The Inverse DWT is used to restore the disparity map to its original resolution.
Hardware Implementation
The entire architecture was designed to be fully pipelined on a Xilinx Virtex-7 FPGA. This design allows data to flow through the stages without stalling, outputting a depth estimate for a pixel on every clock cycle after an initial setup latency. The choice of Haar wavelets is intentional for hardware efficiency, as they can be implemented using simple adders and bit-shifters, avoiding the complex multipliers that consume significant FPGA area and power.
Key Results
The DWT-based approach was evaluated using the Middlebury Benchmark (Tsukuba, Venus, Teddy, and Cones). The results showed dramatic improvements over existing FPGA-based designs:
- Throughput: The system achieved 103 FPS at 1280x720 resolution, representing a 52% increase in frames per second.
- Computational Efficiency: A 44.4% reduction in cycles per frame (CPF) was observed.
- Hardware Savings: Utilization of Look-Up Tables (LUTs) and registers dropped by 61.5% and 59.6%, respectively.
- Accuracy: The system maintained an average accuracy of 90.27%, proving that processing 50% fewer samples results in only a marginal (approx. 3.1%) increase in error compared to full-resolution state-of-the-art designs.
Conclusion
By processing images in the wavelet domain, this methodology proves that high-quality depth maps can be generated with significantly lower computational overhead. The 50% reduction in data workload translates directly into hardware efficiency and speed, making it highly suitable for real-time embedded applications. Future work will explore more sophisticated post-processing to further reduce errors at object edges and validate the system's robustness in diverse outdoor environments.
Summary updated