Continuously evaluating approximate similarity search on high dimensional data stream

In many applications, including online video monitoring system, it is often necessary to find out the most similar sequence, from a sequence set in database, for the high dimensional data stream. Due to its high computation complexity and the restrictions of query processing on data stream, continuously retrieving the similar content on high dimensional data stream is very challenging. The evaluating algorithm is expected to be very fast to match data stream arrival rate. To address this problem, we propose a method called CVNN in this paper. In our method, the sequences in database are represented into the compact summaries, namely CVs, which can be stored in memory. An online algorithm is conducted to transform data stream into a small number of CVs continuously, at the same time the nearest neighbor query is processed periodically based on the similarity of CVs. Experimental results show that our method is fairly effective.