Efficient Extraction of Motion Vectors from H264 Video Streams

Many modern applications use video analysis, and in particular motion analysis, to achieve different purposes, from motion recognition to motion prediction. In this regard, a common interest is to speed up the extraction of this type of data, in the form of motion vectors, from the most commonly used video standard: the H264 Standard. The typical H264 decoders, in particular FFmpeg, decode the entire video to extract the motion vectors - a process that is highly expensive in terms of timings. For this reason, the thesis studies the possibility of building an enlightened version of a decoder composed only of the necessary parts for motion extraction (mainly entropy decoding). The first part of the thesis analyzes this possibility by simplifying the available FFmpeg decoder without optimizing its performance, while the second part deals with the construction of an extractor, following the ITU-T H264 Standard, and the management of motion vectors. The final extractor performs better than the entire decoder, reaching the expectations of an improvement that performs 29% better with the CAVLC and 20% with CABAC entropy coding. Also, as a side effect, the memory usage is largely improved, since the data for the entire frame is not needed anymore. Here, the improvement is even bigger (91% for CAVLC and 78% for CABAC), making this algorithm perfect for simple or embedded systems.