A Comparative Study of High-Speed Deep Learning Frameworks for Real-Time Person Detection in Smart Surveillance Systems

Abstract

Purpose: The goal of the study titled "A Comparative Study of High-Speed Deep Learning Frameworks for Real-Time Person Detection in Smart Surveillance Systems" is to check, Compare, and Examine How well Different top Deep Learning Frameworks Work in Terms of speed, accuracy, and how efficiently they use resources for detecting people in surveillance Settings. The Study wants to find out which Framework is Best at Keeping Detection Accurate While also Using less Computing Power, so it can work smoothly and quickly in Smart Surveillance Systems that have Limited Resources and need to Handle a lot of data at once.

Design/Methodology/Approach: This study looks at two methods, YOLOv11 and SSD, to detect people in real time using images and real surveillance videos. The system was created using Python, OpenCV DNN, and with support for a GPU. To check how well they work, several factors were considered, such as how accurate they are (Precision, Recall, F1-Score, mAP, IoU), how quickly they run (FPS, latency, model size), and how they handle difficult situations like low light, objects blocking the view, and crowded areas. All the tests were done under the same conditions, and the results help decide which method is best for use in smart surveillance systems.

Findings/Results: The comparative analysis revealed that YOLOv11 achieved higher detection accuracy (Precision, Recall, F1-Score, mAP, IoU) with moderate latency, while SSD demonstrated faster inference speed (higher FPS and lower latency) but comparatively lower accuracy, indicating a trade-off where YOLOv11 is more suitable for accuracy-critical surveillance and SSD is better for speed-oriented real-time applications.

Originality/Value: This study compares YOLOv11 and SSD for real-time person detection in smart surveillance systems. It shows how accuracy and speed balance against each other when both models are tested under the same conditions. The research also gives useful guidance to help experts choose the best model for use in real-world surveillance settings.