CHLAC require smaller amount of calculation compared with the other methods for the same purpose. However, high speed CHLAC, that can process more than 30 frames per second in a smaller processing system, has been desired. We have developed a new algorithm that makes CHLAC about 10 times faster than before by parallel computation. Using the algorithm, a small processing system at the notebook PC level can process multiple numbers of camera videos in real time, without relying on a special hardware.
To provide a framework that supports the construction of a system using CHLAC, we developed a platform software “Lavatube” that helps construct a system to process camera videos through the combination of icons. In the past, constructing an image processing system required expertise of image processing and programming. Lavatube, however, enables a developer to easily construct an image processing system by connecting icons. At the same time, he can efficiently develop the system because he can adjust parameters while displaying processing results in real time.
The processing system using CHLAC detects unusual motions in the video by analyzing the feature vectors of camera videos that CHLAC obtains by principal component analysis. Here, CHLAC can detect an unusual motion by defining it as “a motion outside the distribution of usual motions,” without creating a model of an unusual motion that varies with the scene and cannot be defined beforehand. The system learned the sample distribution (subspace) of usual motions beforehand, and the system can detects unusual motions, those are different from the usual motions. In Figure 1, for example, CHLAC memorizes the scenes of opening and closing the locker as usual motions, and precisely detects any other motions (here, prying a locker open).
In addition, CHLAC exhibits an excellent performance even in the open air where the influence of various disturbances such as change in sunlight and shaking of trees is unavoidable. It is possible to eliminate these disturbances, when the system learned such disturbances as parts of usual motions. As shown in Figure 2, the system with CHLAC can detect an unusual motion (overleaping the fence in this case) even in the scene where a pedestrian is found walking among trees swinging in the wind.
Although CHLAC exhibits such an excellent performance, it requires too many calculations to process in real time. To settle this problem, we developed a parallel processing algorithm that allows for real time processing even in processing systems at a notebook PC level. It employs single instruction/multiple data (SIMD) to increase the speed. The SIMD, which is incorporated in the majority of current CPUs, allows a process to be performed on elements in a data set in parallel by using a single instruction. As shown in Figure 3, the speed remarkably increased when an order MMX/SSE2/SSE3 mounted in a Intel x86 processor is employed. With the help of high-speed processing, it has become possible to detect and notify unusual motion at the moment it occurs, instead of after it occurred.
“Lavatube” is a software developed to assist the development of a time-series image processing system using CHLAC etc., and it enables a developer to easily construct a image processing system by connecting icons on a graphical user interface (GUI), as shown in Figure 4. Because he can adjust parameters on GUI, he can concurrently confirm the parameters and the processing results to efficiently construct the system. In addition, inexpensive USB cameras and image files of avi or mpeg formats are supported as standard, and he can use them just by arranging their icons.
As described above, the automation of monitoring cameras has been realized with the help of the high-speed CHLAC that has the ability to detect unusual motions in videos and identify humans, and Lavatube that supports the construction of a image processing system. Though it is an issue that has long been supposed to be difficult to realize despite the strong demand.
> 
