This study proposes LCB-YOLOv8, a video-based real-time surveillance-oriented object detection architecture designed to mitigate safety risks caused by external aerial objects such as drone intrusions, bird strikes, and multiple aircraft approaches in...
This study proposes LCB-YOLOv8, a video-based real-time surveillance-oriented object detection architecture designed to mitigate safety risks caused by external aerial objects such as drone intrusions, bird strikes, and multiple aircraft approaches in low-altitude airspace including airports and vertiports. Low-altitude surveillance imagery is characterized by large homogeneous sky regions, small targets with low contrast, and continuous variations in illumination due to time of day as well as adverse weather conditions such as fog and precipitation. These environmental factors increase false positives and false negatives in general-purpose object detectors such as YOLOv8, thereby degrading operational reliability. To address this limitation, this study structurally redesigns YOLOv8 to better adapt it to the aerial object detection domain, aiming to improve both detection performance and robustness.
LCB-YOLOv8 incorporates three major architectural enhancements. First, an LK-Bottleneck is introduced, which combines large-kernel convolution and depthwise separable convolution in parallel branches to effectively capture global contextual information while suppressing computational overhead. Second, channel and spatial attention mechanisms (CBAM) are integrated into the LK-Bottleneck and C2f blocks, forming LK-CBAM Bottleneck and LK-CBAM C2f, to enhance discriminative target features while suppressing background-like noise. Third, the neck is redesigned using a concat-based BiFPN structure, minimizing information loss in top-down and bottom-up feature flows and improving multi-scale feature preservation and fusion.
Experiments were conducted on a newly constructed aerial object dataset consisting of 36,301 images and 66,276 instances across four classes—Cessna aircraft, helicopters, drones, and birds—collected from both field recordings and public sources. Under identical training conditions (input resolution of 640, 50 epochs, SGD optimizer, and CIoU-based bounding box regression), LCB-YOLOv8 was evaluated against all YOLOv8 scales (n/s/m/l/x). Module-wise ablation studies and performance comparisons across different large-kernel sizes were performed to verify the contribution and design validity of each component.
Experimental results demonstrate that LCB-YOLOv8 consistently outperforms the baseline YOLOv8 across all scales. For the s-scale model, Precision/Recall/mAP@50/mAP@0.5:95 improved from 89.2/88.3/90.6/65.6 to 90.5/89.6/91.7/68.2, with mAP@0.5:95 increasing by up to 2.6 points. Ablation studies confirmed that the large-kernel module is the primary contributor to performance gains, while the combined application of CBAM and BiFPN further improves detection accuracy and localization precision.
These results indicate that the proposed LCB-YOLOv8 can be effectively applied to real-time aerial surveillance and alert systems in low-altitude airspace such as airports and vertiports, where both real-time performance and high reliability are critical.