Social Distancing Detector for Pedestrians Using Deep Learning Algorithms



Jee, Chee Yan (2021) Social Distancing Detector for Pedestrians Using Deep Learning Algorithms. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

[img] Text
Jee Chee Yan.pdf
Restricted to Registered users only

Download (8MB)


Social distancing is the most effective non-pharmaceutical approach to contain the spread of Covid-19. Computer vision technology plays a key role in automating the task of monitoring social distancing measures. YOLOv3 is selected for the social distancing detector due to its realtime and stable accuracy performance. YOLOv3, however, required a powerful GPU computational power to sustain its high accuracy and speed. Such high computational cost may not be beneficial just to maintain its enhancement in accuracy especially in real-world applications where hardware resources may be a constraint. It is very challenging when there are demands on running such an algorithm on a computationally constrained platform to achieve real-time performance. Despite YOLOv3 is the most ideal choice in the real-world environment because its speed and detection accuracy are well balanced, it is still a burden for an embedded device. The algorithm is only suitable for an embedded device if the size of the model is reduced. Hence this is why YOLOv3-Tiny is selected. The selected algorithms are trained and tested with Google’s Open Image Dataset V6. To simulate a real-world situation, The Oxford Town Centre Dataset is selected to test the ability of the trained algorithms. Few hyperparameters are tuned to optimize the performance of the selected algorithms such as the “Early-Stopping” technique, learning rate warmup, Cosine Learning Rate Decay, Data Augmentation technique and transfer learning is applied in this research work. Furthermore, Generalized Intersection over Union (GIoU) is used instead of the conventional method, IoU. Finally, YOLOv3 achieved mAP45 of 54.12% and a speed of 3.04 fps. YOLOv3-Tiny achieved mAP45 of 31.657% and a speed of 14.10 fps. Hence, YOLOv3-Tiny is a preferable choice to become a real-time social distancing detector.

Item Type: Final Year Project
Subjects: Technology > Electrical engineering. Electronics engineering
Faculties: Faculty of Engineering and Technology > Bachelor of Electrical and Electronics Engineering with Honours
Depositing User: Library Staff
Date Deposited: 09 Jul 2021 07:57
Last Modified: 12 Jul 2021 06:32