Automatic Sport Video Mining using a Novel Fusion of Handcrafted Descriptors

Recently, video mining in sports has become an attractive topic for computer vision research. The task of recognizing sports activities comprises two important phases: one focused on feature extraction and another focused on the process of classification. This paper presents a novel version of the discrete co-sine transform (DCT) known as the mean multi-block DCT (MmDCT), which introduces powerful denoising capabilities. Furthermore, we propose a local shape feature descriptor for human activity recognition that combines moment invariants (MIs), the MmDCT, and uniform local binary patterns (ULBPs). The different descriptors utilized have different advantages. For example, MIs are invariant in the presence of linear transformations (rotation, translation, and scaling) and ULBPs are robust against illumination changes. Therefore, these two descriptors provide complementary information. Experimental results on three challenging datasets (KTH, UCFII, and HMDB51) demonstrate that our method can achieve better performance than numerous state-of-the-art methods.