融合CBAM的YOLOv4轻量化检测方法

YOLOv4 Lightweight Detection Method Based on CBAM

基于深度学习的目标检测算法应用于无人机视觉中,会极大提升无人机的场景理解能力,但模型参数量和计算量巨大,难以应用于移动端或嵌入式平台.因此本文提出了一种效果较好的轻量级实时检测模型,采用YOLOv4模型网络作为主要参考模型,使用MobileNet替换主干网络,并通过添加CBAM注意力机制以及Soft-NMS后处理策略来提高模型的准确性.选用PASCAL VOC数据集来测试所提出的轻量级YOLOv4模型,结果显示参数量只有原模型的一半,但速度FPS提升了26.48,精度mAP只下降了0.52%.将所提出的轻量化YOLOv4模型部署Nvidia Jetson TX2低功耗系统以及树莓派上,飞行试验显示在TX2上模型FPS达到了21.8,是原始的YOLOv4的4.74倍,将本算法部署到无人机装载的嵌入式平台上,能够对航拍视野中的车辆目标进行实时识别和定位.

Object classification and detection algorithms based on deep learning will greatly improve understanding capabilities when applied to UAV vision,but the amount of model parameters and calculations are too huge to apply to mobile or embedded platforms.This paper investigated the feasibility and accuracy of a new lightweight real-time detection model with main reference model used the YOLOv4 model network andMobileNet replaced the backbone network,as well as CBAM attention mechanism and Soft-NMS post-processing strategy.The PASCAL VOC data set was utilized to test the proposed lightweight YOLOv4 model.The results showed that the parameter amount was only half of the original model and the speed FPS was increased by 26.48 justwith 0.52%reduced mAP.After the model is deployed on the Nvidia Jetson TX2 low-power system,its FPS has reached 21.8,which is 4.74 times that of the original YOLOv4.The algorithm is applied to the embedded platform of the drone,which benefits to real-time recognition and positioning of vehicle targets in the aerial view.