Person Detection in aerial and UAV images Using Deep Learning Algorithms

Kurzfassung

The application of deep learning methodologies, particularly in conjunction with convolutional neural networks (CNNs), has garnered significant attention within the remote sensing community. Among the primary applications are object detection, image classification, and image segmentation. This report specifically delves into object detection, focusing on the detection of humans. In scenarios like search and rescue operations, it is common practice to survey large areas using downward-facing cameras. However, a notable challenge arises from the discrepancy between training datasets for CNNs, which often feature oblique images, and the nadir aerial images utilized for real-time mapping. To address this challenge, a unique dataset has been curated, comprising solely nadir images captured at varying ground sample distances (GSD) ranging from one to six centimeters. The diversity of the training data is ensured through multiple UAV flights conducted across different locations. The dependency on GSD serves as valuable prior knowledge, amplifying the complexities associated with human detection in aerial imagery. An image depicting a human at one centimeter GSD contains considerably more detailed information compared to the same human depicted in an image captured at three centimeters GSD. This discrepancy underscores the necessity for networks trained across a spectrum of GSDs to reliably detect humans across different scenarios. The problem is further compounded by the small size of humans in these images, typically occupying less than 1% of the image area, making them difficult to detect. Traditional object detection methods struggle with these small objects due to their reduced feature representation after pooling layers in CNNs. To overcome these challenges, this report investigates the use of advanced deep learning techniques, specifically one-stage detectors like the YOLO (You Only Look Once) family, with a focus on the latest YOLOv8 model. YOLOv8 has been selected for its superior balance of speed and accuracy, and its enhancements in feature extraction and detection performance. This investigation includes a comprehensive evaluation of different optimizers and augmentation techniques to improve the detection of small objects in high-resolution aerial imagery. Through this investigation, the report aims to contribute to the advancement of deep learning techniques for object detection in aerial imagery, with potential applications spanning various fields such as security, safety, and humanitarian endeavors.