基于改进YOLOv4的无人机目标检测方法

针对无人机平台由于内存、算力有限而导致检测模型部署困难、检测速度降低的问题,提出了一种基于YOLOv4的改进模型.首先,为了减小模型内存占用、节省计算资源,根据目标尺寸特点,对YOLOv4原模型的预测层进行了改进,将三尺度检测模型改进为双尺度检测模型;其次,对双尺度检测模型进行正常训练,然后将其BN层的缩放因子进行稀疏训练,最后通过裁剪一定比例的通道数以再次减小模型内存占用提升检测速度.实验分析表明,在与原模型检测效果基本一样的情况下,最终改进模型的内存占用减少了60％,仅103 M,FPS提升了35％,达到了58帧/s.

An UAV Target Detection Method Based on Improved YOLOv4

As an application platform in target detection, unmanned aerial vehicles play an incomparable advantage and characteristics in reconnaissance missions. However, the limited memory and computing power of the UAV platform are difficult in detection model deployment and slow at detection speed. To solve the above problems, an improved model based on YOLOv4 is proposed. Firstly, in order to reduce the memory usage of the model and save computing resources, this paper improves the prediction layer of the original YOLOv4 model according to the characteristics of the target size. Secondly, the improved model is trained, and then sparse training and channel pruning on the scaling factor of the BN layer are made to reduce the memory usage of the model again to improve the detection speed. The experimental results show that with the detection results being basically the same, the memory usage of the improved model is reduced by 54%, and the FPS is increased by 35% compared with the original model, reaching 58 frames per second respectively.