Dynamic CNN Parameter Exploration for Multi-Altitude UAV Object Detection

Object detection in Unmanned Aerial Vehicles (UAVs) presents unique challenges due to the variety of altitudes and angles from which images are captured. Traditional Convolutional Neural Networks (CNNs) remain the typical model of choice for object detection, but their high computational demands make them difficult to deploy on UAVs with constrained onboard resources. Additionally, most models generalize within a narrow range of distances, limiting their efficacy at the diverse altitudes that UAVs typically fly at. This study explores the impact of key CNN parameters (input image resolution, network width, and kernel size) on detection accuracy and computational efficiency at different altitudes using two multi-altitude datasets. Our experiments show that the ability to dynamically select specific parameter combinations based on altitude enables efficient utilization of resources (both memory and computational) while maintaining minimal impact on accuracy. The results underscore the significant variation in optimal configurations across altitudes, emphasizing the potential of dynamic CNNs that adapt parameters in real time based on altitude. This research provides actionable insights for optimizing CNN architectures for UAV applications, with implications for efficient deployment in areas such as surveillance, disaster response, and environmental monitoring.