Design and Benchmarking of an Embedded System for Low Power Event Based Vision

Sammanfattning

The ability to perceive the environment with minimal energy is a remarkable characteristic of biological vision. Engineers are still far from achieving comparable efficiency in conventional computer vision, where cameras and processors consume significant power to handle dense frame-based data. Event cameras, inspired by the principles of biological vision, represent an effort to move closer to this efficiency by outputting only pixel-level brightness changes instead of full frames. This results in both high temporal resolution and sparse data that can be exploited for fast, energy-efficient computation. However, the baseline power required to acquire and transmit event data in embedded platforms is still not well characterized, and recent work has emphasized the need for realistic end-to-end benchmarks of complete computer vision pipelines in biologically inspired architectures. To address this gap, this thesis presents the design and benchmarking of a scalable embedded system combining the Prophesee GENX320 event camera with the quad-core Alif Ensemble E7 microcontroller featuring integrated Ethos-U55 hardware accelerators. Results show that the camera consumed less than 3 mW even in high-activity scenes, while the MCU dominated overall system consumption. The findings establish a practical lower bound for eventdriven perception on the studied architecture and provide a reference point for future designs, while also identifying areas where further optimization is possible. Building on this work, future research can utilize the embedded system for comparative benchmarking of complete computer vision algorithms to evaluate power – performance trade-offs.