Minor Object Recognition from Drone Image Sequence

Object detection in drone imagery is an interesting topic in the Computer Vision field. This work was widely applied in traffic analysis and control, rescue systems, smart agriculture, etc. However, many challenges exist in developing and optimizing applications because of object density, multi-scale objects, and blur motion. To partly solve the above problems, this research focuses on improving the performance of the YOLOv5m network based on the advantages of the Bi-directional Feature Pyramid Network (BiFPN), Transformer, and Convolutional Block Attention Module (CBAM). The experiments achieve 68.6% and 42.6% of mAP on the proposed datasets (ISLab-Drone) and VisDrone 2021, respectively. That demonstrates the outperformance of the network comparable to other networks under the same testing conditions.