基于SwinTransformer与动态稀疏卷积的无人机视频目标持续跟踪及检测

Continuous tracking and detection of unmanned aerial vehicle video targets based on Swin Transformer and dynamic sparse convolution

摘要:

在面向非结构空域监测的异质无人机多传感器融合系统中,复杂光电对抗与湍流介质扰动导致成像质量退化,视频流中普遍存在跨模态噪声干扰、低信噪比与谱间混叠等现象,易产生目标表征信息丢失等问题,降低检测精度。为此,提出基于SwinTransformer与动态稀疏卷积的无人机视频目标持续跟踪及检测方法。基于前帧检测置信度动态生成稀疏掩码,实现卷积计算自适应聚焦,抑制背景冗余并增强小目标特征。结合SwinTransformer的全局注意力机制与卡尔曼滤波状态预测,实现目标外观与运动状态的双重建模,提升复杂场景下跟踪连续性,减少目标表征信息丢失问题。引入相似度反馈机制,在跟踪失效时触发补偿检测,增强系统对遮挡、位移等极端场景的适应能力,实现跟踪失效下的快速目标恢复。实验结果表明,该方法即使在身份切换次数为20次的情况下边界框仍能精准包围目标,体现出良好定位精度。在晴天、阴天、傍晚场景下,检测精度(mAP)最大分别为0.96、0.85、0.70,在检测精度方面表现优异,验证了该方法在实际应用中的有效性与场景适应性。

Abstract:

In heterogeneous unmanned aerial vehicle multi-sensor fusion systems for unstructured airspace monitoring, complex optoelectronic countermeasures and turbulent media disturbances lead to degradation of imaging quality, and cross modal noise interference, low signal-to-noise ratio, and spectral aliasing are commonly present in video streams, which can easily result in loss of target characterization information and reduce detection accuracy. Therefore, a method for continuous tracking and detection of unmanned aerial vehicle video targets based on SwinTransformer and dynamic sparse convolution is proposed. Dynamically generate sparse masks based on the confidence level of previous frame detection, achieve adaptive focusing through convolution calculation, suppress background redundancy, and enhance small target features. By combining SwinTransformer"s global attention mechanism with Kalman filter state prediction, dual modeling of target appearance and motion state is achieved, improving tracking continuity in complex scenes and reducing the problem of loss of target representation information. Introducing a similarity feedback mechanism to trigger compensation detection when tracking fails, enhancing the system"s adaptability to extreme scenarios such as occlusion and displacement, and achieving fast target recovery under tracking failure. The experimental results show that this method can accurately surround the target with bounding boxes even after 20 identity switches, demonstrating good positioning accuracy. In sunny, cloudy, and evening scenarios, the maximum detection accuracy (mAP) is 0.96, 0.85, and 0.70, respectively, demonstrating excellent performance in detection accuracy and validating the effectiveness and adaptability of this method in practical applications.