Real-time surveillance system detecting persons in complex scenes

We need a reliable visual surveillance system that can work under various conditions, such as changes of lighting and movements of background objects. In constructing these systems, there are two conditions that must be satisfied. One is that it can distinguish intruders (persons) from clutter in the scene. The other is that it can detect intruders in real time. We propose a method exploiting information over space and time. Firstly, optical flow is computed, and then regions with uniform flow are extracted. The regions are hypothesized as parts of detection targets. Assuming that the targets move in the scene in constant directions at constant speeds, we predict a path of each region and vote it in the space-time cube. If enough votes support a particular path, the method reports the detection of an intruder. We have implemented this method on a real-time image processor which consists of multiple DSP boards and constructed a surveillance system that can detect intruders in real time. We have tested our system in outdoor scenes with various conditions such as shaking trees, a fountain, a rotating signboard and flags. Experimental results show that the system can detect intruders successfully even in such complex scenes. (C) 2001 Academic Press.