Similarity retrieval on video database based upon module operation

Abstract

In this paper, we propose a method of retrieving videos from the video database based upon the temporal relationship among the objects in the video data. We transform each video, Vi, into a set of ordered triples, (Oi,Oj,Rij)s, where Oi and Oj are two symbolic objects and Rij is the temporal relationship between Oi and Oj. Then, we construct a hashing table from all the triples with their corresponding videos in the video database. Every ordered triple is assigned a prime number. Therefore, each video Vi can be transformed into a positive integer Pi. A query can be converted into a positive integer value Pq via the preconstructed hashing table. The answer to a query is a collection of videos, in fact a set of desired frames, such that their corresponding Pi's can be divided by Pq, i.e., the remainder of Pi/Pq is equal to 0. A video database query system is built based upon the proposed method. From complexity analysis, we notice that the query time is with order O(n), where n is the number of videos in the video database, as oppose to O(nm) of the query without module operation, where m is the average number of frames in a video.