Given a video of handwash, this model detects 7 different hand wash steps involved in the video and output the following:
• Display time spent on each activity (hand wash step), and overall time spent in washing hands.
• Displays if any step is missing
• Display the final compliance or non compliance.
Project Aim - To ensure that hand wash is performed in hospitals, homes and in office premises according to WHO norms.
Here is the Kaggle dataset which contains hand wash videos for 7 hand wash steps: https://www.kaggle.com/realtimear/hand-wash-dataset
• To train the model, the dataset is annotated in YOLO format.
• Have a look at the dataset directory structure in HandWashDataset_yoloFormat.
• For each step, the video is first converted into frames.
• 101 frames of 7 classes are annotated (Total Images = 707).
• 81 X 7 (567) frames are used in training set (HandWashDataset_yoloFormat/TrainingData/images/train).
• 20 X 7 (140) frames are used in validation set (HandWashDataset_yoloFormat/TrainingData/images/val).
• Test videos are stored in HandWashDataset_yoloFormat/TestingData.
• The model has been trained using YOLOv5.
The dependencies to install YOLOv5 can be found in (https://github.com/ultralytics/yolov5)
• Our model has been trained on PyTorch framework.
• The image input size is (640,640,3).
• The model is trained for 100 epoch with a batch size of 16.
• The trained model gives validation mAP (mean Average Precision), precision
and recall of 0.996, 0.993, 1, respectively.
| Steps | Images | Labels | P | R | mAP |
|--------|--------|--------|-------|----|-------|
| All | 140 | 140 | 0.993 | 1 | 0.996 |
| Step1 | 140 | 20 | 0.99 | 1 | 0.995 |
| Step2 | 140 | 20 | 0.99 | 1 | 0.996 |
| Step3 | 140 | 20 | 1 | 1 | 0.996 |
| Step4 | 140 | 20 | 0.997 | 1 | 0.995 |
| Step5 | 140 | 20 | 0.991 | 1 | 0.996 |
| Step6 | 140 | 20 | 0.99 | 1 | 0.996 |
| Step7 | 140 | 20 | 0.994 | 1 | 0.996 |
Please take a look at the following colab notebook. It consists of model training and model prediction output.
Colab Notebook:
