Investigates Image data fusion techniques that combine image and track data from multiple sensors to achieve improved accuracies and more specific inferences than could be achieved by the use of a single sensor alone. Our aim is to explore the state-of-the-art image processing algorithms for achieving effective data fusion as in:
1. Introduction
- The goal of intelligent visual surveillance algorithm is to perform surveillance tasks as automatically as possible in real time. Considering that the indoor environment of elevators is usually confined and isolated from the outside, people inside elevators are always at risk of crime. Also, There are various illumination changes depending on the status of the door and lighting conditions and shadows or reflections of humans may be cast on the floor and walls depending on the texture and materials used. For these reasons, developing an intelligent real-time visual surveillance system for elevators is a considerable challenge.
2. Framework of Algorithm
1. Image sequences are obtained from a camera.
2. The blobs of the foregrounds from each frame are segmented by using background subtraction.
3. The motion vectors inside the blobs of the foreground are extracted by optical-flow method.
4. An abnormal event detector involving a frame classifier decides whether the frame is normal or abnormal based on foreground segmentation and motion vector extraction.
3. Foreground Segmentation and Motion Vector Extraction
- Foreground segmentation
1. The foreground segmentation algorithm is a fusion of several techniques including background estimation using GMM, foreground estimation from previous frames, and utilization of other scene information such as motion density and texture component.
2. Assuming that the background includes not only a direct view of the interior but also virtual images of the interior reflected by inner walls.
- Motion vector extraction
1. The optical flow inside the foreground regions between continuous frames is computed in grayscale for motion vector extraction.
2. The Lukas-Kanade method is used for computing the optical flow in the horizontal and vertical directions.
4. Abnormal Event Detection
- Object classification
1. Filter out the foreground blobs, those having sizes too small or being located within positions impossible for a human to be.
2. Classify the foreground blobs into groups or individual objects.
3. Estimate the number of people inside the elevator.
- Abnormal frame detection
1. Find out those frames containing larger average motion magnitudes than the threshold, because fighting between two people involves a lot of movement.
2. Filters out those frames containing a biased direction of motion. Unlike fighting, walking for the purpose of getting on or off the elevator has a clearer direction of motion.
