UAV Farmland Object Tracking Based on Improved QDTrack

In order to further improve the environmental perception capabilities of ultra-low-altitude agricultural drones, an improved multi-objective tracking method based on QDTrack(Quasi-Dense Track) is proposed. We first constructed a video dataset containing typical farmland objects for model training and testing. To improve the detection performance of QDTrack, we used the Mask R-CNN instead of Faster R-CNN as the detection model. The result showed that the MOTA and MOTP value of Mask-QDTrack was 1.5 and 0.3 percentage points higher than the original QDTrack. Expecially, the IDF1 has increased from 20.5% to 41.8%, an improvement of 21.3 percentage points. To address the issue of information transmission between different frames, spatial attention was introduced into the detection network(Mask R-CNN) to focus on the significant apparent features of the tracking target and suppress the influence of useless features such as noise. In addition, deformable convolution was introduced to add the offset, increase the receptive field and improve the robustness of the model. The results showed that the improved-Mask-QDTrack achieved the best MOTA of 63.4%, MOTP of 81.1% and IDF1 of 54%, which was 14.2, 3.3 and 33.5 percentage points higher than the original QDTrack, respectively. And the IDSW of improved-Mask-QDTrack was 8 lower than the original QDTrack. The improved QDTrack multi-object tracking model proposed in this paper can provide technical support for the safe and autonomous flight of UAV.