UAV Based YOLOV-8 Optimization Technique to Detect the Small Size and High Speed Drone in Different Light Conditions

Public safety and other sectors are paying more and more attention to uncrewed aerial vehicle ( UA V) identification as UAVs become more commonplace in commercial and industrial settings. Methods for object detection by UAVs are progressing rapidly, too. Nevertheless, researchers still need to overcome substantial obstacles in this field because of the tiny size of drones, complicated airspace backdrops, and fluctuating light conditions. Addressing these issues, this work suggests a small UAV detection approach that utilizes the improved YOLOv8. The first step is to upgrade the device's recognition capabilities for small marks by addition a high-resolution recognition head According to the study, the YOLOv8 network's large target prediction layer, feature extraction, and fusion layers are eliminated. Only four, eight, or sixteen-sampled feature maps are stored for UAV prediction. the upgraded network architecture, which feeds feature maps from the third C2f layer straight into SPPF for multi-scale feature extraction using 16-times down-sampling. After leaving the Upsample-ConcatC2f module, fused feature maps are immediately attached to the next component. Our approach boosts performance compared to the baseline model by 11.9% for accuracy, 15.2% for recall, and 9% for mean average precision (MAP). The model size is decreased by 57.9 ^% , and the number of limits is decreased by 60%. Also, our approach is better suited for engineering deployment and real-world UAV object recognition system applications, and it shows apparent benefits in comparative studies and experiments using self-built datasets.