Won-Ha Kim

Plume Detection and Tracking System

In this work, we develop a system for tracking a gaseous object. So far, there has been extensive research on segmenting and identifying rigid objects or model-based deformable objects. However, the segmentation and tracking of gaseous objects is rarely studied. In this work, we develop a system for tracking gaseous objects via image segmentation and identification. A good example of such objects is a smoke plume. Plume detection is a specific application for airborne surveillance. Difficulties in tracking gaseous objects arise from the irregularity of objects' shapes, their motion behavior, and the low resolution of the image. Since the intensity of a gaseous object is variable at the boundary, the object's boundaries are not well defined. Furthermore, since the motion of gaseous objects depends on air turbulence, it is hard to model the time evolution of a plume sequence, and the region of a single plume may be spatially disconnected. Therefore, modeling such an object's shape and motion behavior is very intractable. In many practical image sequences, object images are noisy and object sizes are small compared to the image resolution. Thus the texture patterns on plume objects are indiscernible, which makes object identification more difficult.

The above figure illustrates the overall view of the proposed gaseous object tracking system. The motion block matching method detects moving blocks. The spatial clustering clusters the blocks into objects. The temporal clustering tracks motion behavior of the spatially clustered objects and clusters the objects following the same motion path into the same spatio-temporal object. The results of temporal clustering are saved as object memories. The object memories provide object motion behavior for the next temporal clustering process. At a given frame time, the object identification process extracts features of the objects obtained after temporal clustering and identifies the objects that match the desired object features.

The proposed system tracks gaseous objects for which conventional methods such as the optical flow method or model-based methods often fail. It is also more robust to noise and faster than the conventional optical flow method. In addition, the proposed method can be implemented for operational systems via an MPEG-4 code-stream. The developed system has been applied to detect smoke plumes. The system performance was affected by wind strength, plume size, plume volume and motion estimation block size. However, the proposed system can automatically detect most of the plumes that humans can detect by visual inspection. In this work, we assume that the camera platform is fixed and ground based. In real applications, the camera may be moving and jittering. Therefore, we must also develop a system to compensate the camera motion and the jitter noise.

Demonstration (requires AVI movie plug-in; e.g., Windows systems)
Click the .avi file links and a new image window will appear. Click on the image window to run the movie. The purple objects indicate positively identified plumes. The `+' indicates a plume source, that is, a chimney. Blue clusters are plume suspects, i.e., video objects that have been neither identified nor rejected conclusively as plume objects. Yellow blocks are isolated moving blocks that were not rejected as noise but have not yet been clustered and classified as potential plume objects either. All motion blocks are initially classified as yellow in the first frame of each sequence. As the classifier observes more of each video object's spatio-temporal behavior it then promotes each object's classification to blue (potential plume) or purple (definite plume) based on its identification criteria, as detailed in the tech report.
Plume1.avi , Plume2.avi , Plume3.avi , Plume4.avi , Plume5.avi

Tech Report (Postscript file)

For more on video-based plume detection research at Los Alamos, go to the DAPS Project page on plume detection.