Robust Vehicle Detection and Tracking Model via Deep SORT Over Aerial Images
Ghulam Mujtaba, Ahmad Jalal
Abstract
Detecting vehicles is vital for traffic monitoring and surveillance which generally employ cameras on bridges or roadsides. However, aerial imagery offers greater flexibility by covering wider areas with mobile platforms. This study introduces a novel approach for vehicle detection and tracking in aerial image sequences, leveraging advanced image processing and deep learning techniques. Initially, images are preprocessed to reduce noise and enhance brightness using a Gaussian mixture model. For precise vehicle boundary delineation, image segmentation is performed using a Fully Convolutional Network (FCN). The Deep SORT algorithm is then employed for vehicle detection. The Maximally Stable Extremal Regions (MSER) algorithm is applied for vehicle matching to ensure consistent feature identification across frames. Vehicle counting is performed using deep learning-based methods, providing superior accuracy and scalability. Vehicle tracking is executed with the Single Shot MultiBox Detector (SSD) framework, ensuring best performance and high precision. Finally, trajectories are drawn to visualize movement. This integrated pipeline demonstrates significant potential for enhancing real-world traffic management systems. The detection accuracy on UAVDT datasets reached 88%, with a tracking accuracy of 90% which shows superior performance.