Automatic detection of indoor occupancy based on improved YOLOv5 model

Area of Science:

• Computer Vision
• Artificial Intelligence
• Energy Systems

Background:

• Indoor occupancy detection is crucial for energy efficiency and disease traceability.
• Current YOLO algorithms struggle with dense, occluded targets due to anchor-based mechanisms and coupled detection heads.
• This leads to slow convergence and suboptimal performance in real-world scenarios.

Purpose of the Study:

• To develop a novel, efficient, and accurate indoor occupancy detection algorithm.
• To address the limitations of existing YOLO models in detecting occluded individuals.
• To improve energy management and public health monitoring through enhanced occupancy detection.

Main Methods:

• Proposed DFV-YOLOv5: a decoupled, anchor-free convolutional network based on YOLOv5.
• Implemented an anchor-free mechanism to reduce parameters and heuristic tuning.
• Decoupled detection heads to resolve classification-regression conflicts and speed up convergence.
• Utilized VariFocal loss for class imbalance and asymmetric sample weighting.
• Introduced a hybrid activation function (sigmoid-linear unit and rectified linear unit) for improved nonlinear representation and reduced inference time.

Main Results:

• DFV-YOLOv5 demonstrated significantly improved detection accuracy and robustness compared to mainstream models.
• The algorithm achieved shorter inference times, enhancing real-time applicability.
• Experiments on VOC2012, CrowdHuman, and a custom classroom dataset validated the model's performance.
• Ablation studies confirmed the effectiveness of the improved loss function and model design.

Conclusions:

• The proposed DFV-YOLOv5 algorithm offers a practical and effective solution for indoor occupancy detection.
• The anchor-free, decoupled design with VariFocal loss optimizes performance for crowded and occluded environments.
• This advancement supports more efficient energy management and enhanced public health traceability systems.