GICEDCam: A Geospatial Internet of Things Framework for Complex Event Detection in Camera Streams

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Complex Event Detection (CED) is crucial for analyzing camera streams in safety, security, and monitoring.
• Existing CED frameworks struggle with high resource demands, scalability issues, and inaccuracies (false positives/negatives).
• Limited spatiotemporal labels and expensive training hinder the effectiveness of current CED methods.

Purpose of the Study:

• To propose GICEDCAM, a novel framework for efficient and scalable Complex Event Detection.
• To reduce computational cost and end-to-end latency in CED.
• To enhance the accuracy of spatial event detection by minimizing false positives and negatives.

Main Methods:

• GICEDCAM distributes CED processing across edge, stateless, and stateful layers.
• A Spatial Event Corrector component utilizes geospatial data analysis for improved accuracy.
• Evaluation involved 16 camera streams analyzing four complex events against a baseline.

Main Results:

• GICEDCAM achieved a 36% reduction in end-to-end latency and a 45% decrease in total computational cost.
• Performance gains increased with a higher number of objects per frame.
• Bayesian Network (BN) offered lowest latency, Long Short-Term Memory (LSTM) highest accuracy, and trajectory analysis the best trade-off.

Conclusions:

• GICEDCAM offers a scalable and computationally efficient solution for Complex Event Detection.
• The Spatial Event Corrector effectively reduces errors in spatial event detection.
• Different corrector variants provide distinct advantages in accuracy and latency for specific applications.