Electronic Systems with High Energy Efficiency for Embedded Computer Vision

Electronic systems are now widely adopted in everyday use. Moreover, nowadays there is an extensive use of embedded wearable and portable devices from industrial to consumer applications. The growing demand of embedded devices and applications has opened several new research fields due to the need of low power consumption and real time responsiveness. Focusing on this class of devices, computer vision algorithms are a challenging application target. In embedded computer vision hardware and software design have to interact to meet application specific requirements. The focus of this thesis is to study computer vision algorithms for embedded systems. The presented work starts presenting a novel algorithm for an IoT stationary use case targeting a high-end embedded device class, where power can be supplied to the platform through wires. Moreover, further contributions focus on algorithmic design and optimization on low and ultra-low power devices. Solutions are presented to gesture recognition and context change detection for wearable devices, focusing on first person wearable devices (Ego-Centric Vision), with the aim to exploit more constrained systems in terms of available power budget and computational resources. A novel gesture recognition algorithm is presented that improves state of art approaches. We then demonstrate the effectiveness of low resolution images exploitation in context change detection with real world ultra-low power imagers. The last part of the thesis deals with more flexible software models to support multiple applications linked at runtime and executed on Cortex-M device class, supporting critical isolation features typical of virtualization-ready CPUs on low-cost low-power microcontrollers and covering some defects in security and deployment capabilities of current firmwares.