Title, Goal, Objectives, Scope, Expected Performance
FYP Title: A Real-Time Social Distancing Detector Based on Deep Learning Approach
Goal: The monitoring of crowd in R&R based on real-time social distancing using deep learning method. The functionality of this proposed system will be examined through image processing on hardware.
Objectives:
- To develop a real-time social distancing detection system for face mask recognition
- To characterize the performance of detection scheme with input resolution of 416x416 pixels using YOLOv4
- To justify the effectiveness of the real-time social distancing detection system in various frame time and environment
Scope:
Social distancing and wearing face mask have proven to be effective ways to prevent the spread of Covid-19 disease. Hence, this project focuses on developing a real-time social distancing detection system with face mask recognition using YOLOv4 algorithm. The object detection model is constructed in Ubuntu Linux 20.04 to provide platform for Python and OpenCV development. By reducing the video resolution and input resolution, it will greatly increase the speed of the model. Thus, the input resolution of 416x416 pixels is used for YOLOv4 algorithm. In order to analyze the effectiveness of the system, regression of probability analysis is constructed using MATLAB.
Expected Performance:
This project aims to develop a real-time social distancing detection system with face mask detection at R&R to ensure the travellers' safety in this Covid-19 pandemic. It is expected that this system can be developed successfully and able to recognize at high speed and high accuracy. The R&R administrator can use the model results to monitor travellers to avoid infection and enhance their safety.
Figure 1: Flow of Proposed System
Face Mask Detection Testing
Training of Custom Detection Model using YOLOv4-tiny
Figure 1: Flow of Training of Custom Detection Model (face mask)
Based on Figure 1, the training of the face mask detection model is done by the YOLOv4-tiny algorithm. YOLOv4-tiny is the compressed version of YOLOv4. It provides faster training and faster detection. It has only two YOLO heads as opposed to three in YOLOv4 and it has been trained from 29 pre-trained convolutional layers as opposed to YOLOv4 which has been trained from 137 pre-trained convolutional layers. YOLOv4 is more suitable compared to YOLOv4 for real-time object detection as faster inference time is more important than accuracy.
Before dataset training, 1510 images with and without face masks are prepared and the images are then annotated with bounding boxes labeled with or without face masks on people's faces. After that, the values of batch is set to 64 and subdivisions is set to 32 in yolov4-tiny-custom.cfg file for training purpose. Figure 2 shows the mAP (Mean Average Precision) chart for the custom object detection model training. The higher the mAP, the better it is for object detection. The best model produces around 94% of mAP which is good for object detection. The next step is to test the best model with some images, videos and webcam. Figure 3 shows that the model is able to detect person with and without face mask.
Figure 2: mAP (Mean Average Precision) chart
Figure 3 Tested image using the best weights.