多光谱动态融合目标跟踪

Object tracking plays a significant role in various applications, such as activity recognition, video surveillance, human computer interaction, traffic monitor and vehicle navigation and so on. Even though with large progress in object tracking, its performance is suffered from various environmental conditions, such as background, occlusion and illumination etc. Robust object tracking is still a great challenge. In this project, we'll attempt to improve the performance of object tracking with multi-spectral sensors data. The data are fused in a dynamical manner so as to reliably track objects in complex scenes. The object and background show different appearances on the same spectrum. Further, such differences can improve the performance of object tracking. Thus, an adaptive learning is used to simultaneously fuse multi-spectral information and track objects. The information fusion processed in the pixel level is directly used for the task of object tracking and adjusted in term of the tracking result. The information fusion and object tracking promote each other, and are adaptive to the change of the background.

视频跟踪研究课题意义重大，应用极为广泛。比如行为识别、安全监控、人机交互、交通监控和车辆导航等等。尽管目标跟踪研究取得了很大进展，但是，跟踪性能仍然受到各种环境因素的影响，比如背景、遮挡以及光照条件等等。可靠、稳定的目标跟踪仍然是一个很大的挑战。为了克服环境因素对跟踪性能的限制，本项目计划采用多光谱传感器来获取多通道数据，动态融合多光谱信息来自适应跟踪复杂场景中的目标。物体与背景在不同光谱段上成像差异能提高跟踪性能，故而，采用自适应学习机制来实现多光谱融合与物体跟踪同步进行，多光谱图像融合直接服务于物体跟踪，并且根据跟踪的当前结果来指导多光谱融合。多光谱图像在像素层面上根据跟踪的需要进行融合。多光谱图像融合与物体跟踪两者互相影响、互相促进，并随背景变化而动态调整。

视频中目标跟踪是一个极为重要的计算机视觉研究课题，在多个领域中具有较高应用价值。既有重要军事用途，又具有广泛的民事用途。比如，军事目标的瞄准、锁定并跟踪，交通车辆行为检测，车辆导航，人群行为监控等等。本项目通过动态融合多光谱信息，实现目标跟踪的背景自适应。目标与背景在不同的光谱通道上有不同的响应。故而，充分利用这种差异性，将多通道提取的光谱强度特征和纹理特征，较好的结合起来跟踪物体。并在跟踪中，根据背景的变化，调整多光谱信息的融合。目标跟踪成为一个闭环自更新过程。实验结果表明，提出的方法有较好跟踪效果和较强的鲁棒性，能较好的克服光照、遮挡、尺度和背景等因素的影响。