Enhancing Person Detection in UAV Color Images: Empirical Study and Novel Techniques for Improved Performance

This research article presents an in-depth empirical study and optimization approach for enhancing person detection in UAV (Unmanned Aerial Vehicle) color images. The study focuses on fine-tuning the state-of-the-art object detection algorithms, mainly Faster R-CNN, to maximize its performance for detecting individuals in UAV imagery. Key aspects such as dataset utilization, backbone architecture selection, loss function choices, and post-processing approaches are thoroughly investigated in this work. The study illustrates the integration of the novel feature upsampling operator CARAFE into the feature pyramid networks of pre-existing object detection models. Furthermore, it uses bounded IOU loss function for training and soft-NMS post-processing techniques to improve the detection performance for small objects like persons. The research also explores SAHI (Slicing Aided Hyper Inference) during training and inference to enhance the detection of small objects in aerial images. Combining the Manipal-UAV dataset with standard VisDrone and Okutama datasets, the study creates a comprehensive training dataset, leading to improved generalization. Extensive evaluations under varying conditions provide valuable insights, making this research significantly contribute to advancing person detection in UAV imagery for critical applications like search and rescue, surveillance, and security.