An Automated Object Detection and Tracking System in Surveillance Aerial Images Using 3D Adaptive Multi‐Dilated Yolov8 With Attention Mechanism

ABSTRACT Nowadays, the need of detecting targets and performing tracking in aerial images using drones along with power sensor devices has increased. Both drone technology and object detection models provide a major contribution to agricultural irrigation, battlefield reconnaissance, traffic management, and forest patrolling. These kinds of tasks are accomplished with superior accuracy by drone technology. But, the prior techniques faces several challenging problem due to wider image perspective, and enormous tiny objects in the image. Various prior deep learning mechanisms show poor resolution in handling the tiny samples. These kinds of outcomes contribute to poor localization accuracy in the network architecture and also the image fractions are settled with the objects. In order to mitigate these issues, the research leverages a new deep learning‐based object detection and tracking approach using surveillance aerial images. Initially, the required surveillance aerial images are gathered by benchmark resources. Further, the garnered images are transferred into the 3D Adaptive Multi‐Dilated Yolov8 with Attention Mechanism (3DAMDY‐AM) for detecting and tracking the objects in surveillance aerial images. The proposed Improved Lyrebird Optimization Algorithm (ILOA) is employed for optimizing the parameters. At last, the numerical experiments are conducted for the developed model, contrasted with the existing related techniques. The accuracy outcomes of the recommended technique show 95.14%. To ensure better statistical analysis, the best measure of the developed model shows 1.74%, 1.44%, 4.58%, and 1.73% better performance than MTBO‐3DAMDY‐AM, OOA‐3DAMDY‐AM, POA‐3DAMDY‐AM, and LOA‐3DAMDY‐AM. The mean Average Precision (mAP) measure of the developed approach shows 74.19%, 36.11%, 77.35%, and 80.92% better performance than CNN, N‐FINDR, Yolov8, and 3DAMDY‐AM, respectively, at the batch size of 8. The significant advancement in the recommended technique guarantees to enhance the system's efficacy and reliability. In summary, the developed model shows a promising outcome, but the work needs to expand on validating real‐time data. Moreover, it needs to validate dynamic scalability conditions with diverse surveillance applications in overcrowded scenarios. Further, the multi‐data source is also required to improve the model performance on diverse factors like scene times, weather conditions, and locations, respectively.