Improved Object Detection and Recognition for Aerial Surveillance: Vehicle Detection with Deep Mask R-CNN

Object detection and recognition methods are critical in aerial surveillance applications, allowing us to detect and recognize objects through unmanned aerial vehicles. However, most approaches rely on deleting the background to achieve accurate detection. Current object detection methods in aerial imagery, such as principal component analysis, suffer from low accuracy and high computational load. Therefore, accurate vehicle detection in aerial images is essential for various applications, including traffic nursing, urban planning, and disaster response, facilitating decision-making processes in these areas. This study aims to detect vehicles in aerial images using the Deep Mask-R-CNN deep learning framework. The dataset of aerial sensing images used in this study is from DOTA. The pre-trained VAID dataset is fine-tuned to make the model more suitable for aerial image-specific characteristics. The results indicate that the proposed method performs exceptionally in detecting automobiles from aerial photographs, with high accuracy, low false positives, and high recall rates. The Deep Mask R-CNN framework is well-suited for this task as it provides object detection and segmentation capabilities. It can be useful in applications where vehicle shape and size are important. Additionally, the proposed method is robust in handling scale variations and occlusions, making it suitable for day-to-day scenarios. Furthermore, the model can be easily integrated with other computer vision algorithms, such as multi-scale feature extraction and object tracking, for further improvements.