计算机视觉
计算机科学
人工智能
接头(建筑物)
目标检测
跟踪(教育)
对象(语法)
雷达跟踪器
视频跟踪
特征提取
噪音(视频)
估计
噪声测量
图像分割
跟踪系统
模式识别(心理学)
传感器融合
作者
Yuanxun Wu,Dawei Zhang,Yanchao Wang,Yabin Wang,Chengzhuan Yang,Xiaowei He,Zhonglong Zheng
标识
DOI:10.1109/jsen.2026.3678754
摘要
With the widespread application of unmanned aerial vehicle (UAV) in fields such as environmental monitoring and urban planning, vision-based multiple object tracking (MOT) technology has become crucial for dynamic scene perception. UAV scenarios present critical challenges for MOT, including high-altitude viewpoints, significant scale variations, and frequent occlusions. To address these issues, this paper proposes StrongDepthMOT, an innovative joint detection and depth estimation framework that integrates object detection, depth estimation, and drone motion pose estimation to significantly improve tracking performance. The framework first introduces a dynamic upsampling strategy to adaptively adjust sampling positions and weight coefficients, reducing upsampling artifacts while preserving spatial details. Then, frequency-based depth decoder is designed to dynamically balance high- and low-frequency information, improving contour sharpness and texture clarity in depth maps. In addition, our channel attention fusion module employs deformable convolution for cross-layer alignment, dynamically recalibrating multi-scale features to enhance detection accuracy. Experiments demonstrate that StrongDepthMOT achieves state-of-the-art performance on mainstream UAV datasets such as VisDrone2019 and UAVDT, significantly reducing identity switches (IDs) while improving robustness in occluded and multi-scale scenarios. Visualization results further confirm its effectiveness in depth estimation and trajectory association. This work provides an optimized depth-aware solution for UAV-based MOT technology in aerial remote sensing.
科研通智能强力驱动
Strongly Powered by AbleSci AI