Compressed domain video zoom motion analysis utilizing CURL

This paper explores the application of the concept of CURL borrowed from vector calculus to the zoom motion detection and classification problems. The interframe block motion vectors extracted from the compressed bitstream form the input to the proposed method. These block motion vectors are analyzed by partitioning the motion vector field into 4 representative quadrants followed by quantizing the block motion vectors into 3 levels and converting the block motion vectors into complex motion vector space. The resultant vector for each of the 4 quadrants is estimated followed by estimating the velocity vector between the quadrants. The CURL of the velocity field is then estimated whose magnitude essentially provides the area enclosed between the resultant quadrant motion vectors which are utilized for separating the zooming and non-zooming camera types. The zooming camera frames are further classified into zoom-in and zoom-out types utilizing the direction information (anti-clockwise/clockwise) extracted from CURL of the velocity field. The novelty here stems from the fact that a concept borrowed from vector calculus is being applied to the zoom motion analysis problem. Although handcrafted features from CURL are utilized we demonstrate its superiority over existing methods including a deep learning architecture where we show the robustness of the proposed features extracted from CURL in the presence of noise. Experimental validation carried out utilizing block motion vectors extracted using Exhaustive Search Motion Estimation algorithm as well as H.264 decoded block motion vectors demonstrate superior performance for the proposed method both in terms of detection accuracy as well as computational complexity in comparison to existing techniques.

Automated summary

This paper explores the application of the concept of CURL borrowed from vector calculus to the zoom motion detection and classification problems. The proposed method utilizes interframe block motion vectors extracted from compressed bitstream and analyzes them by partitioning the motion vector field, quantizing the block motion vectors, and converting them into complex motion vector space. The estimated CURL provides information about the enclosed area between quadrant motion vectors, which is used for separating zooming and non-zooming camera types. Zooming camera frames are further classified into zoom-in and zoom-out types based on the direction information extracted from the CURL of the velocity field. Experimental validation shows the superiority of the proposed features extracted from CURL.