Object-Camera Motion Classification Using Semantic Features and Neural Network on Sports Video

The presence of object-camera motion makes automatic content-based sports video analysis (CBSVA) a challenging task. The three main kinds of object-camera motion which often occur in sports videos are large object movement, Camera panning motion, and Camera zoom in or out motion. Contents that carry similar features as these three motions are responsible for increasing false detection and miss classification rates in CBSVA tasks. Hence, classifying object-Camera motion using its semantic features is highly needed. Existing methods of content-based video analysis do not use individual semantical features of these motions. Therefore, this paper proposes a novel method of object-camera motion classification by segmenting foreground and background of the video frames using Convolutional Neural Network firstly, secondly analyzing pixelbased visual semantical motion features from foreground and background using optical flow velocity, and lastly classifying video frames into three classes of object-camera motion.