Deep Learning & Video Analytics

Share             

AI-based Video Analytics for Pandemic Management



 
Background
 
In response to the COVID-19 pandemic, the Singapore government has introduced stringent measures to enforce everyone wearing a face mask and keep safe distancing in public areas. The crowd and personnel flow control have reached unprecedented levels of importance in order to flatten the curve. More security officers and enforcement officers have to be deployed to patrol around the streets, HDB, and food centers to ensure safe-distancing and compliance with government policies. In order to relieve the stress of our police force and increase the efficiency of the patrolling activities, an intelligent and pro-active surveillance and crowd control system is needed.
Based on the successful deployment of NTU ROSE Lab’s bespoke person searching and person re-identification AI system at foreign worker isolation facilities in the Changi Exhibition Center, we developed a more generic visual analytics platform for rapid deployment on any surveillance systems. By utilizing various AI models developed by NTU ROSE lab, the proposed platform can perform quick person searches and contact tracing within the surveillance system. It also proactively monitors the area by raising alarms on the social distancing violation and detecting people without a face mask. The proposed visual analytics platform enables 24x7 monitoring of streets, malls, hotspots, and any other facilities and delivers actionable intelligence with in-depth insights to combat the spread of the COVID-19.
 
 
Face Mask Detection
In the ROSE pandemic management system, we use the state-of-the-art face detection targeting the very small faces in the surveillance system. The face-mask detector will detect whether the person is wearing a mask or not. If it’s negative, the system will trigger an alarm. For a properly worn mask, the mask must cover a person’s nose and mouth. Figure 1 shows the one of the alarms when someone is not wearing a mask and labelled by the red bounding boxes.
 
 
tumbnail_action_recognition.png 
Figure 1: Detection of people not wearing masks properly.
 

 
Social Distancing Analyzer
The social distancing analyzer in the ROSE pandemic management system can automatically analyze the distance between people. When people are too close to each other, it will automatically trigger an alarm. Figure 2 shows the analyzer can detect people which are close to each other (red bounding boxes) and also differentiate people who keep a reasonable distance to others (green bounding box). Deploying it on current surveillance systems and drones to monitor large areas can help to prevent the spread of the coronavirus by allowing automated and better tracking of activities happening in the area. The analyzer provides analytics of the area in real time. It can also be used to alert security personnel in case of considerable violation of social distancing protocols in a particular area.
 
 
tumbnail_action_recognition.png 
Figure 2: Social Distancing Measuring.
 

 
Population Density Heat Map for Crowd Control
The ROSE pandemic management system can actively detect the number of people under each camera. This will give us the estimated crowd density near each camera cluster. The crowd density heatmap of each floor can be generated automatically for better visualization, as shown in Figure 3. It will provide valuable information for surveillance officers for a more efficient patrolling force deployment in the crowded areas.
 
 
tumbnail_action_recognition.png 
Figure 3: Real-Time Crowd Density Heat Map for each Floor
 

 
Person Searching and Retrieval
The person searching and retrieval functions are also well integrated into the system for better contract tracing purposes. The contact tracking can be achieved by two main functions: trajectory tracking retrieval and real-time person matching. The trajectory tracking retrieval aims to find the person of interest (POI) in all cameras and plot the historical movement trajectory of the POI within a building or campus, as shown in Figure 4. The real-time person matching aims to match the POI in real-time surveillance cameras and raise the warning to the surveillance officers, shown in Figure 5.
 
 
tumbnail_action_recognition.png 
Figure 4: Trajectory tracking retrieval example.
 
 
 
tumbnail_action_recognition.png 
Figure 5: Real-time matching example.
 

 

 
​​​​
​​​
 

 Sample Videos

 
Face Mask Detection Video Demo: ​​​
 



Social Distancing Video Demo:
Retrieval Demo: ​​​
 



Matching Demo:
​​