Intruder Detection From Video Surveillance Using Deep Learning

• Published in: 2024 IEEE 15TH CONTROL AND SYSTEM GRADUATE RESEARCH COLLOQUIUM, ICSGRC 2024
• Main Authors: Abu Mangshor, Nur Nabilah; Sabri, Nurbaity; Aminuddin, Raihah; Rashid, Nor Aimuni Md; Johari, Nur Farahin Mohd; Jemani, Muhammad Adib Zaini
• Format: Proceedings Paper
• Language: English
• Published: IEEE 2024
• Subjects: Automation & Control Systems; Engineering
• Online Access: https://www-webofscience-com.uitm.idm.oclc.org/wos/woscc/full-record/WOS:001345150000016

Video surveillance, or closed-circuit television (CCTV), is a well-known technology globally. Many homeowners use this technology for security purposes. However, the existing system cannot distinguish whether the individual captured in the footage is the homeowner or a stranger. Moreover, the authority still needs to analyze each footage, which can be time-consuming. Therefore, this present study introduces a system that can distinguish the homeowner from any stranger, record a specific chunk of footage when an intruder is detected, and notify the homeowner about the incident. To ensure the success of this project, two Deep Learning (DL) models were trained: the EfficientDet and MobileNets models. The first model, EfficientDet is an object detection model used for the detection of a person. The second model, MobileNets is the image classification model for performing figure recognition of the homeowner. These deep learning models are loaded into a Raspberry Pi 4 (Pi) to act as video surveillance and perform detection together with classification. If an intruder is detected, a notification will be sent to the homeowner together with a short video recording of the incident which is viewable via a web application. Based on the testing performed, the system passed all use cases of the functionality testing. This study confirmed the usability as well as the accuracy of the proposed technique. On accuracy testing, the object detection model achieved an average precision (AP) of 76.00%. As for the image classification model, the accuracy achieved is 85.71%. The ability of this system can be improved with the usage of better hardware for inference such as Google's Coral, an edge TPU, as this would increase the performance of the model in terms speed.