Intelligent vision system for computing at the edge
Edge computing involves processing information locally where it is generated, and where the results of computation will be used. It is needed to make smart devices truly smart, i.e. not reliant exclusively on feedback from a remote server. “Edge computing enables systems to work autonomously, and to rapidly react and adapt to changes in their surrounding environment,” explains MISEL(opens in new window) project coordinator Jacek Flak from the VTT Technical Research Centre(opens in new window) of Finland. “They can operate even when network connection is unavailable – for example in unpopulated areas or zones of natural disasters – and avoid many security- and privacy-related issues induced by data transfers.”
Compact, efficient artificial vision system
To be commercially attractive and applicable across numerous applications however, smart devices need to be efficient and reliable. The EU-funded MISEL project sought to address this by focusing on the performance of the entire processing chain, from sensors providing input data all the way to final decision-making. The project gathered researchers from various research fields, including material science, electronics and algorithm design. “Our main objective was to develop a compact artificial vision system able to perceive and interpret changes in the environment,” says Flak. “For this, we worked on a number of aspects of an artificial vision system.” These included bio-inspired cellular sensors and processors designed to mimic certain functionalities of the eye’s retina and certain parts of the brain. An image sensor was also developed to adapt to different lighting conditions. A layer of quantum dot-based photodetectors was integrated to extend sensitivity beyond visible light towards the near-infrared spectrum. “We also introduced novel algorithms for processing visual information,” notes Flak. “We tried to find the best possible trade-off between complexity and performance, to allow for efficient implementation in small, low-powered devices.”
Successful hardware-algorithm co-design
Hardware capabilities were extensively studied during algorithm development, providing everyone with a better understanding of the possibilities and challenges of achieving efficient and reliable edge computing. The outcome of this was a successful hardware-algorithm co-design. “Our work also opened up new possibilities for efficient data representation of image analysis and hardware implementation,” remarks Flak. “On the manufacturing side, we identified numerous technical solutions for better and more reliable devices. The designed electronic circuits, though very complex, all tied together and worked as designed.” The designed and evaluated systems are also full of edge-AI accelerators – special processing cores dedicated to solving complex operations required by AI algorithms.
New environmental perception systems
The MISEL project helped to resolve some key communication and computing bottlenecks in edge computing and also revealed improvement needs for next versions. “With the knowledge obtained from the project, it will be easier to target future products towards autonomous mobile robotics applications,” adds Flak. For example, based on the results of MISEL, new environmental perception systems can be developed to target autonomous robots or drones. This could be very useful in practical applications such as the search for survivors in disaster zones, security monitoring or future smart car needs such as assistance in collision avoidance. “Our project partner Kovilta(opens in new window) has announced plans to further develop the tested accelerator structures in the field of autonomous mobile robotics, including drones and automotive industry,” says Flak. “Other consortium partners including VTT are expanding their services, thanks to the technologies developed or improved within the project.”
