FEATURE Machine Vision
From vision to perception on low-power computing platforms
RoboK CEO, Hao Zheng, discusses advances to make computer vision with deep learning widely adopted
(left and right) Monocular depth estimation in an automotive application
W
ithin a traditional machine- vision approach, a key step for image classifi cation is to identify an object
using feature extraction. Techniques such as edge and corner detection are used to gather important information from the raw image. The results are then compared to pre-defi ned specifi cations of diff erent classes of the objects to fi nd a match. For the same image classifi cation task, under a deep-learning approach, a model will be trained using a large amount of the data to ‘learn’ what to look for regarding each specifi c object class, unlocking the underlying patterns and developing an AI-based inferencing system. The benefi t is that this system is continually learning and improving the quality of its identifi cation.
Machine vision
Whilst traditional machine-vision approaches can be suffi cient and cost- eff ective for tasks such as defect detection on a production line, they lack in areas such as obtaining spatial information. For example, when recovering 3D information, a stereo camera is often needed, and conventional image processing techniques could use triangulation to estimate depth information. However, not only is recovering absolute scale using images from a monocular camera diffi cult, but, also, gaining relative scale often relies on determining the structure from motion, and performance deteriorates in situations where there is motion blur. Deep-learning techniques have the potential to address these physical and
18 December/January 2021 | Automation
inherent limitations and remove any bottlenecks, but require a large amount of data for training to gain high accuracy. So, there are costs associated with data collection, labelling and graphics processing. Also, deep learning may not be able to be generalise, so retraining models is often needed, which is cost prohibitive. Therefore, a trade-off between robustness/ accuracy and cost effi ciency is a key deciding factor when choosing the most appropriate approach.
Addressing the dilemma RoboK has addressed this dilemma by reducing the computation time required. It has developing a new method for fusing raw data directly from sensors, such as cameras, radars, GPS and IMUs, and for performing depth estimation to gain 3D information. This computer-vision approach is supported by a novel and highly-optimised AI-based perception module to enable intelligent insights to be gained rapidly and effi ciently, which is vital for fast decision-making, for example in advanced driver assistance systems (ADAS). This has all been achieved while signifi cantly reducing the memory and computing requirements, to enable it to run on low-power computing platforms. Although RoboK’s technology is designed for the automotive environment, increasingly the design and implementation of ADAS is achieved within a simulation environment, and a low-power approach also enables a cost- eff ective creation of faster simulations. This benefi t has been recognised by Siemens
Digital Industries Software, which selected RoboK to support the development of a closed-loop simulation within PAVE360 – a pre-silicon autonomous validation environment – capable of testing an entire vehicle with an unlimited number of complex driving simulations. Simulation for the entire system
requires simulation of software (including simulated sensors input), underlying operating systems, electronic systems and all the mechanical components of a physical vehicle. It is highly time- consuming to run a high-fi delity test- driving scenario. With this validation platform, system designers can guarantee that each software- and hardware-design iteration can be tested and validated virtually, quickly and, most importantly, before any hardware is produced.
The speed and intelligence off ered for ADAS also makes the RoboK’s computer- vision solution suitable for industrial applications, such as collision avoidance for moving robots and industrial vehicles on the factory fl oor, as well as quality assurance for the production line. As warehouses and factories become
increasingly automated and connected, AI-based computer vision that could deliver robust performance at the edge will improve the aff ordability and accelerate the adoption of such powerful technology.
https://robok.ai CONTACT:
RoboK
automationmagazine.co.uk
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
