Building an Intelligent Aerial Surveillance System through Deep-learning based Object Detection

The article discusses the challenges involved with detection of vehicles with aerial pictures and methods used to overcome these challenges. Aerial images present unique challenges, such as smaller size of vehicles & intricate backgrounds. Traditional approaches such as sliding window and feature extraction have limitations in accurately detecting vehicles in aerial images. Deep learning models like R-CNN, Faster R-CNN, & Mask R-CNN were proposed & demonstrated exceptional performance. However, there remain challenges in their direct application to vehicle detection in aerial images. Modifications were proposed to overcome these challenges, including improvements to region projected network & classifier and the use of instance segmentation in addition to object detection using bounding boxes. The article highlights the need for large aerial picture real-time detection and the importance of accurate vehicle recognition from aerial photographs in various applications, including urban planning and traffic management.