Evaluation of tracking in video sequences

Observation of long sequences of video images in surveillance applications may encounter several problems due to camera motion or rotation, unexpected size and speed of objects, variation of color due to sunshine and shadowy areas. Robust tracking algorithms are needed to compensate for the variations of different recording conditions. in this paper we evaluate the detection probability of our tracking algorithm with ROC curves and with synthetic degradation methods. Recorded experimental multi-sensor data is used to compare the accuracy in different spectral bands. Moving object detection in a guarded area can produce many false alarms due to the moving environment such as trees and bushes, birds and animals. By applying tracking and classification, false alarms can be reduced avoiding unnecessary recordings and preventing the displacement of guards. Track speed, size, direction and range (distance to camera) are calculated. The objects are classified roughly into classes as person, vehicle, and fast moving object or simply as moving object. The results of the algorithm applied to the experimental data and the algorithm evaluation are presented.