Published: Aug 14, 2021

shaping our future with computer vision technology

With its proven viability and potential in various applications across the globe, research in computer vision has since advanced at a whirlwind pace. At this level of focus, we have seen a vast number of breakthroughs in terms of improving existing technology or adjusting its suitability to specific use cases – the key to customisation being to build a successful computer vision application based on the model.

Our playbook at NCS adopts a practical approach, where new technology such as transfer learning to modify an existing model is utilised, instead of building a model from scratch – allowing for efficient customisation and deployment for each unique scenario within a short span of time.

With specific requirements from different customers, existing pre-trained models may originate from data sets of distinct characteristics – greatly benefiting a specific application while utilising the customer’s data. Concentrating further research and refinement of existing, or creation of new technology will see it perform a broader range of functions.

Utilising proprietary Deep Learning technology, NCS has been able to offer one such solution. Theorising a potential infection by people who cough and sneeze, the AI has been able to pick up behaviours such as head movement during a sneeze or an arm covering a face, especially when the person has coughed or sneezed for a prolonged period. Thereafter, the location is detected in real-time, of which a moving patrol robot sends an alert to a command centre.

• Highly customisable to identify actions of interest along with localisation of all individuals involved
• Video based approach with a novel time persistence algorithm to confirm the action, reducing false alarms
• Light-weight solution that supports all three major deployment topologies - AI Edge device, on-prem and Cloud servers
• Near real-time detection - within a second of occurrence on edge devices
• Edge variant is PDPA compliant with no image or video data sent to servers
• Automatic face masking ensures privacy of individuals

State-of-the-art edge analytics technologies such as Deep Learning, Object Detection, Wide-angled cameras, and Millimeter-wave sensors provide additional assistance to regulating social distancing in mass numbers, sounding an alert upon non-adherence. Such edge solutions run on compact and modern AI platforms which consume very little power.  

• Light-weight solution that supports all three major deployment topologies - AI Edge device, on-prem and Cloud servers
• MMwave sensor data does not reveal individual identities – not being image data – allowing for PDPA compliance
• Customisable in detecting different types of human postures such as standing, sitting and falling
• Near real-time detection - within a second of occurrence on edge devices

Here, we tackle the problem of patient falls in hospitals – a common but calamitous occurrence in hospital care, particularly elderly patients. More than one-third of falls in hospitals lead to injury or serious trauma, leading to additional medical bills or further health complications. In one of Singapore’s public hospitals, NCS is currently running a proof-of-concept – a demonstration of feasibility - to address this, which aims to detect, understand, and analyse patient movement behaviour when leaving a bed. Using state-of-the-art deep learning methodology, movement behaviour is analysed through time to detect and prevent a fall. This approach is performed on an AI edge device, and the patient’s face is automatically censored to protect his/her privacy.

• A first in analysis of individual body part movements to predict intentions to exit a bed
• Patients’ intentions include analysis of surroundings to form a prediction
• Light-weight solution that supports all three major deployment topologies - AI Edge device, on-prem and Cloud servers
• Near real-time fall detection - detection within a second of occurrence on edge devices
• Edge variant is PDPA compliant with no image or video data sent to servers

Automatic face masking ensures privacy of individuals

In discussion with various local agencies on proof-of-concept opportunities – we utilise a technology that taps into existing surveillance cameras to detect fights that affect public safety – with wide-ranging utility in areas like prisons, nightlife hotspots, or schools. Employing a lightweight deep learning method, human body movements such as punching or kicking for a consistent period, are learnt to detect a fight - which thereafter automatically alerts the relevant authorities and is deployed on an AI edge device for privacy protection.

1. Highly customisable to identify actions which constitute aggression - such as punches or kicks – and can localise the individuals involved  
2. Video based approach with a novel time persistence algorithm to confirm aggression – reducing false alarms
3. Light-weight solution that supports all three major deployment topologies - AI Edge device, on-prem and Cloud servers
4. Near real-time aggression detection within a second of occurrence on edge devices
5. Edge variant is PDPA compliant with no image or video data sent to servers
6. Automatic face masking ensures privacy of individuals

To accurately identify, and track movement and interactions in public spaces in a non-intrusive way for purposes of enhancing contact tracing serves the public in a pandemic situation – with a potential solution utilising computer vision algorithms to analyse video footage from existing CCTV cameras, identifying the person of interest and others in close contact.

Traditional facial recognition technologies may suffer from the inability to identify faces when the facial images from surveillance cameras are of low quality, or if people are masked. Using our efficient person re-identification technology, the machine learning model is trained by consideration of various human face and body features to achieve accurate detection capability, across different cameras. Our lightweight solution can be deployed on multiple platforms according to specific requirements.

• State-of-the-art solution to perform recognition and re-identification based on full body regions
• Recognises and automatically identifies people who were in proximity – which is potentially used to identify the chain of people who came into contact with a specific individual
• Light-weight solution that supports all three major deployment topologies - AI Edge device, on-prem and Cloud servers
• Near real-time Person Re-Identification detection - within a second of occurrence on edge devices

Edge variant is PDPA compliant with no image or video data sent to servers

Our research focuses on real-world problems and solutions – contributing to the betterment of society. The recent pandemic was a proper test, which drove our focus toward meeting new challenges that were otherwise never experienced. However, with current technology and tools available to us, adaptation and evolution are still required to face ever-changing market demands.

Most of our work is geared towards the detection of events (Reactive approach), with the next step being the prediction of such events (Proactive approach), such as workplace safety or human intention. Similarly, non-invasive sensors will be explored further to address the privacy concerns of the general public. Data fusion across different sensory streams to better understand people or environments are another direction that we are interested in and will be explored in due course.