Detecting Human Actions in Drone Images Using YoloV5 and Stochastic Gradient Boosting

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Human action recognition from Unmanned Aerial Vehicles (UAVs) is crucial for applications like environmental monitoring and search and rescue.
• Challenges include variable scales, orientations, occlusion, and processing constraints inherent in drone-based image acquisition.
• Existing methods struggle with the complexities of real-world aerial data.

Purpose of the Study:

• To develop and evaluate a low-resource machine learning (ML) method for human action recognition from UAVs.
• To investigate the effectiveness of combining object recognition with classifier techniques for single-image action identification.
• To assess the proposed method's performance on the real-world Okutama-Action dataset.

Main Methods:

• Utilized the Okutama-Action dataset, which captures representative action scenarios under controlled image acquisition parameters.
• Developed an architecture integrating YoloV5 for scalable and efficient object recognition.
• Employed a gradient boosting classifier to handle samples of variable difficulty, complementing YoloV5.
• Conducted an ablation study to test different YoloV5 architectures and evaluate the combined approach.

Main Results:

• The proposed YoloV5 and gradient boosting pipeline demonstrated superior performance compared to previous architectures on the Okutama-Action dataset.
• The enhanced performance is attributed to the efficiency of YoloV5 and the pipeline's suitability for dataset specificities.
• The approach effectively balances bias-variance tradeoff for aerial action recognition.

Conclusions:

• The integrated YoloV5 and gradient boosting approach offers an effective low-resource solution for human action recognition from UAVs.
• This method shows promise for real-world applications requiring robust action identification in challenging aerial environments.
• The findings highlight the importance of tailored pipelines for optimizing performance on specific datasets.