Visual Surveillance on DSP-based Embedded Systems

Inffeldgasse 16a/2nd floor, A-8010 Graz, Austria

Visual surveillance has become an important topic of research in the last few years due to the increased need for security in public places and the unbowed trend to use digital cameras for surveillance purposes, or for integration of visual sensors in other objects of personal use, like mobile phones or PDAs. To perform dedicated operations at camera site, embedded platforms, so-called \emph{Smart Cameras}, have become popular recently. These platforms are equipped with one or multiple video sensors and enough computational power to process the data stream onboard. Moreover, these platforms have additional advantages, like robustness against environmental stress or low power consumption. The use of smart sensors facilitates the building of large networks, and the local processing paradigm allows for extracting valuable information at site and transmission over low-cost and low-bandwidth communication channels.

However, until now, little work has been done on the development and investigation of state-of-the-art algorithms for surveillance from the field of computer vision in respect of embedded systems. On this account, in this thesis we focus on the tasks of object detection and object recognition on smart cameras. We discuss several issues of algorithm development on embedded DSP-based platforms, especially the issues related to parallelism mechanisms and fixed-point arithmetic. Given an extensive overview about current work on object detection and object recognition, we investigate both fields of development in detail. After discussing the basic algorithm, we experimentally demonstrate the suitability of the approach, derived in this thesis, for performing object detection in real-time on a real-world traffic surveillance scenario, given a prototypical DSP-based hardware platform. Our results encourage the integration of our algorithm into a larger system for public surveillance. In the context of object recognition, we demonstrate, how state-of-the-art recognition technology can be used to deploy reasonable recognition capabilities on smart cameras. After proving the suitability on a moderate size object database, we show how our algorithm can be used in a traffic surveillance scenario for recognition and reacquisition of vehicles on public streets. The main advantages of our approach are the high accuracy and the minimization in communication necessary between adjacent camera motes. By investigating our approach in detail, we are able to draw general conclusions and statements about the suitability of different aspects and methods of software development on DSP-based embedded systems.

PDF