Dimensionality Reduction with Random Projection and Distance Space for Video Similarity Measurement: Application with Sports Video Classification

This paper proposes the video similarity measurement approach for sports video classification by dimensionality reduction with random projection (RP) and distance space. Most video data are huge files, which vary in terms of length and amount of data, resulting in time-consuming data processing; therefore, reducing the dimensionality of the data becomes a necessity. All frames of training videos are extracted by color histogram based method. After that, all features of videos are projected onto a lowdimensional subspace by RP for reducing the dimensionality of the data. Afterwards, the clustering technique is performed to provide the centroids of each cluster, called reference vectors. Distance from each reference vector in database to the observation sequence is distance space which is the new feature space. Finally, videos will be classified by term weighting and the nearest neighbor classifier. Accordingly, the proposed approach helps enhance feature dimension reduction, resulting in faster data processing. The experimental results show that the proposed approach outperforms the other approaches significantly in sports video similarity measurement.