Professor Xu Tingfa's research team from the School of Optoelectronics at the Beijing Institute of Technology (BIT) recently achieved a breakthrough in the field of multispectral target tracking.

Addressing the bottlenecks of traditional RGB vision in complex aerial scenes, such as difficulties in detecting small targets, severe occlusion, and texture feature degradation, the research team developed the world's first multispectral multi-object tracking (MMOT) dataset from a drone perspective, and proposed an integrated multispectral target perception tracking method.

The related findings were published under the title MMOT: The First Challenging Benchmark for Drone-based Multispectral Multi-Object Tracking in NeurIPS 2025 (Conference on Neural Information Processing Systems), a top-tier international academic conference in the field of artificial intelligence and machine learning (CCF-A category).

The MMOT dataset developed by the team covers various real-world scenarios such as urban neighborhoods, transportation hubs, and sports venues, including 125 video sequences and over 480,000 high-precision annotated target boxes, spanning eight typical categories (such as pedestrians, cars, and bicycles). The multispectral imaging ranges from visible to near-infrared, encompassing a variety of challenging scenarios such as extremely small targets, dense targets, image blur, severe occlusion, fast motion, and in-plane rotation.

The release of MMOT fills a gap in the field of multispectral target tracking for drones both domestically and internationally, providing researchers with the first unified benchmark and open-source framework. This work breaks through the traditional reliance on spatial features in visual perception by fully leveraging the differences in the spectral dimension. It lays an important foundation for future cross-modal fusion perception and the recognition and tracking of intelligent unmanned systems in complex environments.

Paper details：Tianhao Li, Tingfa Xu, Ying Wang, Haolin Qin, Xu Lin and Jianan Li. MMOT: The First Challenging Benchmark for Drone-based Multispectral Multi-Object Tracking [J]. Advances in Neural Information Processing Systems, 2025