History will remember 2020 as COVID and unless something is done, the same can be said for 2021. My team and I hope that by making use of image recognition software we can reduce the manpower strain on ensuring safety guidelines and measures by providing insightful data through filtering. For example, CCTV footages of places where people frequently do not wear masks or of large gathering can be better targeted to ensure that people are following the safety measures in place.
What it does
VidRecog is a machine learning model that is able to detect if humans are wearing their masks through a static image or a real-time video. VidRecog allows uploading of media files (images/videos) to ensure that people are wearing their masks and if the model picks up that someone is not wearing a mask it will flag the file and a human can verify the authenticity of the file and take appropriate actions.
How we built it
We built upon an existing image recognition model that is able to identify objects and we re-trained the model to detect humans with masks. The main underlying technology used is darknet, an open-source neural network in C, however instead of using darknet, we found a similar library, tensorflow-yolov4 which allowed us to create and train models. Additionally, we used labelmg to label our images with bounding boxes in yolo format which can then be used to train our model.
Challenges we ran into
Figuring what kind of object detection/image classification model we want to use, eg. YOLO vs R-CNN. Finding a proper dataset which we can train our model with. Making sure the dataset is of the correct format such that the model can use it to train.
Accomplishments that we're proud of
The accuracy of our model is close to 80%.
What we learned
We learnt how to create and train a custom machine learning model, mining for data/files that we need to use to train our model, data labelling and filtering of noisy image files that might affect accuracy of the trained model. As well as a simple frontend for the user-experience.
What's next for VidRecog
VidRecog The image recognition model can be further improved to yield more accurate results by providing larger and more diversified datasets.
For the application, we foresee that it can be used as an automation tool which can help filter media files for a certain image or object detection so that the user would not have to look through all the media files but only high probabilistic ones picked out by the model. For example, if someone is committing a crime or in covid-19 case when a person is not wearing a mask.
Log in or sign up for Devpost to join the conversation.