A Real-Time Video Detection Algorithm for UAV Based on YOLOv8

Aiming at the common problem of small target detection in UAV real-time video surveillance, this paper proposes a UAV video detection algorithm based on improved YOLOv8 framework. By embedding a feature enhancement module specifically for small targets in the original YOLOv8 architecture, the algorithm effectively improves the recognition degree of the network for small targets. In order to avoid the excessive expansion of network parameters, the header network is simplified, the convolutional layer is removed and the number of channels is reduced to maintain the computational efficiency of the algorithm. Then this paper introduces the Channel Space Attention Mechanism module (CSAM), which makes the network pay more attention to the small target-dense region on the feature map by optimizing the weight distribution, thus enhancing the pertinence of target detection. In addition, a location-based label assignment strategy (LB-SimOTA) is proposed, which assigns different weights according to the intersection ratio (IoU) between the prediction frame and the real frame to optimize the overall quality of the network prediction frame. Experimental results on the VisDrone2019 dataset containing many small targets show that compared with the benchmark model YOLOv8-S, the proposed algorithm improves the accuracy of vehicle and pedestrian detection tasks by 8.44% and 8.12%, respectively, and achieves a frame rate of up to 90 FPS while maintaining the detection speed.