Image Processing Architectures for Binary Morphology and Labeling

Conventional surveillance systems are omnipresent and most are still based on analog techniques. Migrating to the digital domain grants access to the world of digital image processing enabling automation of such systems, which means extracting information from the image stream without human interaction. The resolution, frame rates, and functionality in these systems are continuously increasing alongside the number of video streams to be processed. The sum of all these parameters imposes high data rates and memory bandwidths which are impossible to handle in pure software solutions. Therefore, accelerating key operations and complex repetitive calculations in dedicated hardware architectures is crucial to sustain real-time performance in future advanced high resolution and frame rate systems.

To achieve this goal, this thesis presents four architectures of hardware accelerators to be used in real-time embedded image processing systems, implemented as an FPGA or ASIC. Two morphological architectures performing binary erosion or dilation, with low complexity and low memory requirement, have been developed. One supports static, and the other locally adaptive flat rectangular structuring elements of arbitrary size. Furthermore, a high-throughput architecture calculating the distance transform has also been developed. This architecture supports either the city-block or chessboard distance metric and is based on adding the result of parallel erosions. The fourth architecture performs connected component labeling based on contour tracing and supports feature extraction. A modified version of the morphological architecture supporting static structuring elements, as well as the labeling architecture, has been successfully integrated into a prototype of an automated digital surveillance system for which implementation aspects are presented. The system has been implemented and is running on an FPGA based development board using a CMOS sensor for image acquisition. The prototype currently has segmentation, filtering, and labeling accelerated in hardware, and additional image processing performed in software running on an embedded processor.