A modular CUDA-based framework for scale-space feature detection in video streams

Multi-scale image processing techniques enable extraction of features where the size of a feature is either unknown or changing, but the requirement to process image data at multiple scale levels imposes a substantial computational load. This paper describes the architecture and emerging results from the implementation of a GPGPU-accelerated scale-space feature detection framework for video processing. A discrete scale-space representation is generated for image frames within a video stream, and multi-scale feature detection metrics are applied to detect ridges and Gaussian blobs at video frame rates. A modular structure is adopted, in which common feature extraction tasks such as non-maximum suppression and local extrema search may be reused across a variety of feature detectors. Extraction of ridge and blob features is achieved at faster than 15 frames per second on video sequences from a machine vision system, utilizing an NVIDIA GTX 480 graphics card. By design, the framework is easily extended to additional feature classes through the inclusion of feature metrics to be applied to the scale-space representation, and using common post-processing modules to reduce the required CPU workload. The framework is scalable across multiple and more capable GPUs, and enables previously intractable image processing at video frame rates using commodity computational hardware.