Scene-Specialized Multitarget Detector with an SMC-PHD Filter and a YOLO Network

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• You Only Look Once (YOLO) is an efficient object detection network.
• YOLO performance degrades with significant variations between training and real-world data.
• Automatic customization of YOLO for diverse datasets is challenging.

Purpose of the Study:

• To develop a novel transfer learning algorithm for automatic YOLO network customization.
• To enhance YOLO's robustness against data variations and improve detection accuracy in real-world scenarios.
• To enable automatic labeling of unlabeled target sequences for YOLO retraining.

Main Methods:

• Proposed a transfer learning framework integrating Sequential Monte Carlo Probability Hypothesis Density (SMC-PHD) and Gaussian Mixture Probability Hypothesis Density (GM-PHD) filters.
• Implemented automatic frame labeling for unlabeled target sequences.
• Utilized detection probability and clutter density from PHD filters to retrain YOLO for occluded targets and clutter.
• Developed a novel likelihood density incorporating YOLO confidence and visual context for target sample selection.
• Introduced a resampling strategy for SMC-PHD YOLO to mitigate weight degeneracy.

Main Results:

• The proposed framework successfully automates YOLO network customization using unlabeled data.
• Automatic labeling of target sequence frames was achieved.
• Retrained YOLO demonstrated improved performance in detecting occluded targets and handling clutter.
• The novel likelihood density effectively selected target samples.
• The resampling strategy addressed the weight degeneracy problem in SMC-PHD YOLO.
• Experimental results showed positive outcomes compared to state-of-the-art frameworks.

Conclusions:

• The developed transfer learning framework effectively customizes YOLO networks for improved real-world performance.
• The integration of SMC-PHD and GM-PHD filters offers a robust solution for handling data variations and challenging detection scenarios.
• The method provides a significant advancement in automatic object detection customization.