Detecting Objects in Surveillance Videos with Deep Neural Networks for Crime Scene Analysis
I. Sudha, G. Kirubasri, M. Deivakani, Vijendra Pratap Singh, T. Rajesh Kumar, Mudligiriyappa Niranjanamurthy · 2023
Surveillance systems are critical components of modern security and forensic investigation, and their efficacy is strongly reliant on precise object detection in video recordings. This study looks into the use of deep neural networks to improve object detection in surveillance films for crime scene analysis. This study investigates the capabilities of cutting-edge deep learning architectures in recognizing and classifying objects in various surveillance contexts. The study’s findings are extensive, including detection accuracy metrics for numerous item classes such as “Person,” “Vehicle,” and “Suspicious Item.” Precision values range from 0.78 to 0.92, recall values range from 0.82 to 0.90, and F1 scores range from 0.80 to 0.90, demonstrating the models’ ability to recognize objects accurately, but with variances among item categories. This study also looks into computational performance, offering information about inference times and GPU utilization. Inference times for ResNet-50 and YOLOv3 are 15 ms and 20 ms, respectively, with GPU use percentages of 75% and 90%. These findings provide useful information for picking models that fulfill real-time processing needs while optimizing computational resources. The research also examines the connection between video resolution, detection speed (up to 30 frames per second), and average detection accuracy. Lower resolutions allow for faster processing, but at the expense of accuracy, whilst higher resolutions provide finer details at the expense of larger computational needs. These trade-offs are critical considerations when building surveillance systems to meet certain operational requirements.