Training of new face mask detection model
Some improvements are done:
- The face mask model includes 3 classes: with_mask, without_mask, incorrect_mask _wearing with a total number of 3800 images.
- The best mAP achieved is 94.83% (> 90%) and average loss of 1.3972. It is considered great result for real-time detection.
- However, there might be some error detection on images but still consider acceptable (e.g: incorrect_mask_wearing is detected as with_mask)
mAP and loss chart
Testing Image 1
Testing Image 2
Testing Image 3
Testing Video
Social Distancing Detection Model
- Capture first frame of the input video as an image (Figure 1)
- Select 4 points to be transformed into bird's view, 4 points are connected with lines. (Figure 2)
- Transform the captured first frame image into bird's view. (Figure 3)
- Then fed in again the input video, and transform it into bird's view using the selected 4 points. (Figure 4)
- The blue line shows the selected region to be used to detect person. (Figure 5)
To be improved:
- For now, the whole region of the video is detected (not the selected region).
- Use the Figure 4 for person detection.
- Check for violation distance.
Figure 1
First Frame of Input Video
Figure 2
Chosen 4 points connected in line
Figure 3
Image in bird's view
FIgure 4
Sample frame of input video that has transformed into bird's view
FIgure 5
Person detection on original video