Masked AI: AI Solution for Face Mask Identification

Masked AI is a cloud-based AI solution for real-time surveillance that keeps an eye on the human who violates the rule by not using face masks in public places.

Comment 1

Inspiration

The total number of Coronavirus cases is 5,104,902 worldwide (Source: World o Meters). The cases are increasing day by day and the curve is not ready to flatten, that’s really sad!! Right now the virus is in the community-transmission stage and taking preventive measures is the only option to flatten the curve. Face Masks Are Crucial Now in the Battle Against COVID-19 to stop community-based transmission. But we are humans and lazy by nature. We are not used to wear masks when we go out in public places. One of the biggest challenges is “People not wearing masks at public places and violating the order issued by the government or local administration.” That is the main reason, we built this solution to monitor people in public places by Drones, CCTVs, IP cameras, etc, and detect people with or without face masks. Police and officials are working day and night but manual surveillance is not enough to identify people who are violating rules & regulations. Our objective was to create a solution that provides less human-based surveillance to detect people who are not using masks in public places. An automated AI system can reduce the manual investigations.

What it does

Masked AI is a real-time video analytics solution for human surveillance and face mask identification. Our main feature is to identify people with masks that are advised by the government. Our solution is easy to deploy in Drones and CCTVs to “see that really matters” in this pandemic situation of the Novel Coronavirus. It has the following features:

1. Human Detection

2. Face Masks Identification (N95, Surgical, and Cloth-based Masks)

3. Identify human with or without mask in real-time

4. Count people each second of the frame

5. Generate alarm to the local authority if not using a mask (Soon in video demo)

It runs entirely on the cloud and does detection in real-time with analysis using graphs.

How we built it

Our solution is built using the following major technologies:

1. Deep Learning and Computer Vision

2. Cloud Services (Azure in this case)

3. Microservices (Flask in this case)

4. JavaScript for the frontend features

5. Embedded technologies

I will be breaking the complete solution into the following steps:

1. Data Preparation: We collected more than 1000 good quality images of multiple classes of face masks (N95, Surgical, Clothe-based masks). We then performed data-preprocessing and labeled all the images using labeling tools and generated PASCAL VOC and JSON after the labeling.

2. Model Preparation: We used one of the famous deep learning-based object detection algorithm “YOLO V-3” for our task. Using darknet and Yolo v-3, we trained the model from scratch on 16GB RAM and Tesla K80 powered GPU machine. It took 10 hours to train the model. We saved the model for deploying our solution to the various platforms.

3. Deployment: After training the model, we built the frontend which is totally client-based using JavaScript and microservice “Flask”. Rather than saving the input videos to our server, we are sending our AI to the client’s place and using Microsoft Azure for the deployment. We are having on-premise and cloud solutions prepared. At the moment, we are on a trail so we can’t provide the link URL.

After building the AI part and frontend, We integrated our solution to the IP and CCTV cameras available in our house and checked the performance of our solution. Our solution works in real-time on video footage with very good accuracy and performance.

Challenges we ran into

There are always a few challenges when you innovate something new. The biggest challenge is “The Novel Coronavirus” itself. For that reason, we can’t go outside the home for the hardware and embedded parts. We are working virtually to build innovative solutions but as of now, we are having very limited resources. We can’t go outside to buy hardware components or IP & CCTV cameras. One more challenge we faced was that we were not able to validate our solution with drones in the early days due to the lockdown but after taking permission from the officials that problem was not a deal anymore.

Accomplishments that we're proud of

Good work brings the appreciation and recognition. We have submitted our research paper in several conferences and international journals (Waiting for the publication). After developing the basic proof-of-concept, We went on to the local government officials and submitted our proposal for a trial to check our solution for better surveillance because the lockdown is near to be lifted. Our team is also participating in several hackathons and tech event virtually to showcase our work.

What we learned

Learning is a continuous process. We mainly work with the AI domain and not with the Drones. The most important thing about this project was “Learning new things”. We learned how to integrate “Masked AI” into Drones and deploy our solution to the cloud. We added embedded skills in our profile and now exploring more features on that part. The other learning part was to take our proof of concept to the local administration for trails. All these “Government Procedures” like writing Research Proposal, Meeting with the Officials, etc was for the first time and we learned several protocols to work with the government.

What's next for Masked AI: Masks Detection and Recognition

We are looking forward to collaborating with local administrative and the government to integrate our solution for drone-based surveillance (that’s currently in trend to monitor internal areas of the cities). Parallel, The improvement of model is the main priority and we are adding “Action Recognition” and “Object Detection” features in our existing solution for even robust and better solution so decision-makers can make ethical decisions as because surveillance using Deep Learning algorithms are always risky (bias and error in judgments).

Share this project:

Updates