LUMF-YOLO: a lightweight object detection network integrating UAV motion features

In recent years, the network structure has become more complex to improve the accuracy of convolutional neural networks (CNN), increasing computing power requirements. However, the edge computing capability of unmanned aerial vehicles (UAVs) is low, which makes it challenging to meet the computing power requirements of large neural networks. Therefore, achieving real-time and accurate multi-object detection on UAVs with limited computing resources becomes more challenging. To solve this problem, we present LUMF-YOLO, a lightweight object detection network for UAVs with limited computing power. We employ MobileNetV2 as the backbone, replacing CSPDarknet53 to reduce parameter count and enhance inference speed. We integrate PWC-Net for optical flow estimation and fuse its features with the first, third, and sixth layers of the feature pyramid to improve small object detection. We remove the layer with the largest receptive field from YOLOv4 to focus on detecting smaller targets. An adaptive copy-paste method is introduced to expand the dataset for small and underrepresented objects. LUMF-YOLO achieves 43.3% mAP@0.5 on the VisDrone dataset, demonstrating superior accuracy and performance compared to existing lightweight networks.