With 6.31M confirmed cases under the attack of COVID-19 in India. Maintaining sanity and hygiene remains the only way to prevent and fight the spreading of this deadly virus and making every individual aware of their sanitary health and ongoing COVID-19 epidemic.
Wearing masks is a necessity as per medical practitioners whenever you are at public places or places with high probability of people gathering, to reduce transmission rates. Manual monitoring would require a huge manpower engagement to overcome these we propose a real-monitoring of people wearing medical masks and determing the regions with high transmission chances.
-
- The model is trained on a custom dataset of 10000 images (with over 90000 labels)
-
- Streams videos after being run through detection model of different locations.
-
- Camera ID based line graph of number of people wearing masks to number of people not wearing masks on the basis of periodic frame counts.
-
- Camera ID based line graph of ratio (of nomasks to masks) to time.
-
- The analytics provided about the detection can be exported as .avi files.
- CCTV feed - The model is provided with CCTV video feeds for detection.
- Detection Procedure - The model detects masks and no masks from the CCTV feeds.
- Ratio Calculation - Calculates the ratio of number of no-mask to mask in each frame and a finally its average across a periodic frame count is used to store in a .csv file.
- Graphical Representation- Two line graphs are plotted on the basis of the .csv files.
- mask to no-mask
- ratio to time
- Remote Detection Streams- The CCTV feeds are streamed with detections.
- Overlay Warning- The loactions with ratio lower than threshold value are marked with higher transmission chances
- Operating System
- Linux (with xterm)
- Deep Learning
- Darknet Framework
- CUDA 10.1
- CUDNN
- OpenCV(with CUDA support)
- APIs
- Python Script
- Front-end
- PyQt5
With time to increase the accuracy of the model we wish to enrich our dataset further more.
We wish to incoporate unsupervised learning to reduce periodic updates of the model for upgrading its functioning precision.
We wish to develop an algorithm to identify a person occuring in multiple frames as one to overcome the redundancy in the ratio calculation and its plot.
Biparnak Roy | Debojit Ghosh | Subhadip Nandy |