DESIGN AND IMPLEMENTATION OF SMART SURVEILLANCE SYSTEM WITH YOLOV7 ALGORITHM

• Main Author: Johan Syah Djula, Edi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75315
• Institution: Institut Teknologi Bandung

The Smart Surveillance System, also known as Smart CCTV (Closed Circuit Television) in this study, is an evolution of the traditional Smart CCTV concept that incorporates deep learning technology, a subfield of machine learning, to process data extracted from Computer Vision for real-time detection and tracking of vehicle objects. This system functions as a standalone device, requiring no centralised processing or supporting video image processing applications. Unlike traditional Smart CCTV, which transmits moving images or videos for processing in centralised devices or supporting applications, data generated by vehicle classification is processed directly, providing output in the form of text data, resulting in less data transmission. The researchers used three YOLOv7 variants, namely YOLOv7, YOLOv7-Tiny, and YOLOv7-E6, with DeepSORT as the object tracking system. The results show that the YOLOv7-Tiny model outperforms the YOLOv7 model by 96.23% when compared to YOLOv7 and 97.22% when compared to YOLOv7-E6. Furthermore, the YOLOv7-Tiny model achieved an inference time per frame of 0.0235 seconds, which was faster than the YOLOv7 (0.6226 seconds per frame) and YOLOv7-E6 (0.8449 seconds per frame) models. Memory usage was reduced by 52.94% during the optimisation process, which involved switching the desktop environment from GNOME to LXDE. The study also included Object Tracking and Counter System testing with varying Max Age parameters, with Max Age 40 achieving the highest accuracy level of 100% for inbound objects and 98.03% for outbound objects. The implementation of the YOLOv7 and DeepSORT models in the Smart Surveillance System has proven to be effective in real-time detection and tracking of vehicle objects. The findings of the study show the potential application of machine learning and deep learning technologies in the development of intelligent surveillance systems for traffic monitoring and security that are responsive and efficient. Furthermore, the system generates output in the form of text data, allowing for lightweight and efficient data transmission.