FEATURE SENSORS
Neil Huntingdon, CSO at CronAI explains why it’s time to rethink perception strategies to meet the real world needs of machines
T
o create safe and accurate systems that are truly autonomous and
connect seamlessly with the world around us, we need to design machines and processes that can perceive the environment as we do. Perception strategies - including the
choice of sensor(s), off -the-shelf compute hardware and vendor-specific or 3rd party perception software algorithms are locked down at the design phase. This approach constrains future development for additional or changing applications as well as the level of intelligent perception developed by the machine over time. What’s more, data used to develop
perception algorithms at the design stage tends to be captured for a limited range of scenarios or contexts. Such a siloed approach means that perception systems architectures cannot support real time adaptation of software and computing engines in the field. As a consequence, they are ill-equipped to deal with the unknowns of real time situations where contexts are constantly changing and where the perception strategy needs to dynamically adapt accordingly. Ultimately, these traditional strategies
have non-scalable, non-adaptable and competing computing architectures that were not designed to process the next generation of algorithms, deep learning and artificial intelligence required for 3D sensing mixed workloads. Furthermore, the available computing
architectures for users looking to develop or deploy perception systems tend to be generic and designed for either graphic rendering or data processing, leaving solution providers little choice but to promote these options.
36 DECEMBER/JANUARY 2021 | ELECTRONICS
BRINGING FORWARD A NEW APPROACH 3D sensing is advancing and evolving, and this is driving a shift from cloud computing to edge computing. There are several reasons for this: achieving reduced latency, managing unpredictable network bandwidth and enabling autonomous and continuous local processing for mission critical applications. This shift brings its own unique set of
nuances, such as constraints on the computational resources available, power and energy consumption by edge processing, the memory requirements of algorithms and AI and the design area of the processing platform. This means that engineers need to strike a delicate balance between accuracy and design constraints. To help with this, we can use TOPs
(trillions of operations per second) to measure the theoretical processing power that hardware can provide. TOPs is measured for chips by calculating the number of MAC hardware units and their frequency. For example, 1000 MACs running at 1Gigahertz (1ns per MAC) equals 1 TOPS. It is, however, worth noting that when
processing architecture claims a certain number of TOPs as its processing capability, the theoretical maximum processing capability of the architecture can be overlooked, given that the workload and data path match the processing architecture and its memory hierarchy. So, when workloads don’t match the requirement of the computing architecture, these TOPs are never actually achieved in practice. High TOPs has traditionally been
equated with better performance. However, it is our belief that by designing a platform that delivers high
Building the autonomous future
throughput (frames processed per second) and low latency (delay) even with low TOPs, users can get the performance they need with a lower design area, lower power consumption and lower costs. Our aim therefore is not to provide
higher processing, but an intelligent way to use lower processing to achieve higher throughput with high accuracy. Overall, 3D sensing data processing is more unique and exhibits deeply pipelined sequential, highly parallel as well as mixed workloads. It is this view that has led our
development of senseEDGE - a contextually aware 3D sensing perception processing platform built from the ground up to bridge the gap between complex raw data produced by 3D sensors and a huge range of end user applications. From intelligent transport solutions and smart spaces through to security and surveillance via industrial and warehouse automation, this platform provides the ability to perceive the world the way we do as humans, and take actions accordingly. The FPGA-based inference edge
platform is built around a real-time scalable and adaptable computing architecture that’s flexible enough for algorithms and software to scale and adapt to different workloads and contexts. Its real time contextual awareness means that at any given time the entire edge platform is aware of its external context, the sensor and sensor architecture and the requirements of the user application. AI and automation is set to be truly
everywhere in our future. Our refreshed vision for 3D data processing design is how we see the future of perception.
CronAI
www.cronai.ai/.com / ELECTRONICS
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46
