ACFM: adaptive channel weighted fusion algorithm for improving small object detection performance in UAV traffic

TL;DR

This paper introduces ACFM, a new algorithm that improves small object detection in drone traffic by enhancing feature representation and reducing background interference.

Contribution

The novel ACFM module combines multi-scale refinement, sparse attention, and adaptive weighting for better small object detection in UAV traffic.

Findings

ACFM improves localization consistency and detail preservation of small objects across different resolutions.

The method achieves a maximum mAP gain of 0.8% and 1.3% on VisDrone2019 and UAVDT datasets.

ACFM remains robust in complex scenarios, showing a 0.5% mAP improvement on the AU - AIR dataset.

Abstract

In terms of small objects in drone traffic monitoring, problems like insufficient feature representation, serious background interference, and poor multi-scale adaptability are often encountered. Especially when dealing with complex traffic situations, poor context linking between objects as well as poor detection in congested regions are more noticeable. In order to solve the above problems, we put forward an adaptive channel weighted fusion module, which is ACFM. First, we build a multi-scale refinement module, which can do cross-scale feature interaction via a downsampling-upsampling path. It is combined with a residual calibration mechanism to greatly improve both the localization consistency and the detail preservation of small objects over different resolutions of feature maps. And then, a grouped sparse mask attention module is created to reduce background noise through channel grouping and sparse gating techniques in order to enhance the local saliency features of sparse small targets. Finally, we add a channel-wise adaptive weighting by using the global context. Using an α weight generator which can change the contribution of features based on scene complexity, it get rid of traditional fixed combination plans. From the experimental results, we can see that adopting GFL as the detector, ACFM achieves better performance improvements on the widely used VisDrone2019, UAVDT dataset, and the maximum gain in mAP is more than 0.8% and 1.3% higher than the comparison methods. On the AU - AIR dataset, ACFM is still a little bit better than the other 0. 5% mAP, it is still robust in the complicated situation.

Figures