Automatic Video Detection And Identification Of Legacy Items

Dropped object detection and recognition in video is an important research topic in the fields of artificial intelligence and pattern recognition. Unattended dropped objects in public cause personal loss and damage of public security because the real-time video cannot be detected and recognized automatically. So in actual life, it is not only very helpful to detect and recognize the dropped objects from surveillance video, but also to find and track the pedestrians who own the dropped objects for the purpose of reducing personal loss and anti-terrorism when the dropped objects appear in public, which also meets the current social demand.In terms of dropped object detection, an algorithm based on regional information is proposed to detect dropped objects and analyse relevancy. The algorithm uses foreground detection method based on dual-direction background modeling, MeanShift tracking method and pixel-based regional information at the drop-off point for dropped object detection and relevancy analysis. The algorithm is focused on the division event instead of classifying all the blobs detected. So it performs dropped object detection and relevancy analysis more quickly than the existing methods and meets the need of real-time monitoring better. The experimental results demonstrate the rapidity and accuracy of the algorithm.In terms of dropped object recognition, a dropped object recognition algorithm based on moment invariant and PCA (Principal Component Analysis) is proposed. The algorithm firstly constructs geometrical affine invariant moment for detected regional image of dropped objects, than uses PCA to reduce its matrix dimension, finally uses the decision rule of minimum distance classifier to perform object image matching for the purpose of dropped object recognition. Because the algorithm combines the features of geometrical affine invariant moment and PCA, it can solve the problems of angle rotation, scale variation and so on which occur in the actual dropped objects better. The algorithm can accurately recognize the dropped objects in video and improve the recognition speed. It can process the data of real time video streams better. The experimental results demonstrate that the algorithm can recognize dropped objects in video quickly and accurately and it is appropriate for processing the data of real time video streams.Because the centralized video surveillance model has the limitation of processing large-scale real time video streams, the algorithms of dropped object detection and recognition are realized in a distributed video surveillance model.