Newer AI model stumbles in the dark, older one holds steady
YOLOv8, the latest computer vision system for detecting intruders, performs worse than its predecessor in low light and bad weather—findings that could force security companies to reconsider their technology rollouts. The research suggests one-size-fits-all AI deployments in surveillance systems carry hidden risks.
Originaltitel: Real-time efficiency of YOLOv5 and YOLOv8 in human intrusion detection across diverse environments and recommendation
Intrusion Detection Systems (IDS) are essential for securing areas such as industrial and construction sites. However, when implementing IDS as a service, confidence scores (confidence) provided by YOLOv8 are the most reliable metric as compared to the YOLOv5 available to take appropriate actions to secure these sites and prevent intruders. However, prior research has focused on YOLO's human detection capabilities (whether it can detect or not), neglecting real-time performance in IDS. To address this gap, we propose and present comparative analysis of YOLOv5 and YOLOv8 in a real-time across diverse environmental conditions (luminance, indoor/outdoor, simulated weather). Our findings reveal an average performance of YOLOv5 (outdoor: 90.5%, indoor: 79.1%), YOLOv8 (outdoor: 99.1%, Indoor: 77.2%) confidence in real-time, with a logarithmic relationship between luminance and confidence. Outdoor environments perform better then indoor for both YOLOv5 and YOLOv8, while adverse weather conditions significantly reduce YOLOv8’s effectiveness and increase the efficiency of YOLOv5. Therefore, this enables IDS integrators to adjust minimum confidence thresholds to minimize the risk of preventing potential intruders. However, the consistent and inconsistent confidence scores by both YOLOv8 and YOLOv5 respectively, and impact of weather remains inconclusive due to simulated fog.