Frame rate up-conversion of real-time high-definition remote surveillance video

Abstract

Many surveillance and monitoring systems capture video using static cameras with relatively low frame rate. Low frame rates have benefits such as less storage requirements and less bandwidth used. These issues may be of even greater concern for high-definition video capture and remote surveillance. Also, a reduced frame rate makes the video appear less smooth, causing increased strain for the viewer. It is therefore desirable to increase the frame rate without increasing bandwidth use or changing equipment. This thesis presents the design and implementation of a tool for performing frame rate up-conversion in real-time. The tool makes use of GPUs and multi-core CPUs found in modern computer systems. Techniques and frameworks such as OpenMP, SSE and OpenCL are utilized to make full use of the systems capabilities. Frame rate up-conversion is performed in two steps: motion estimation and motion compensation. For motion estimation a number of block matching algorithms were evaluated and implemented. We present in this thesis our version of an exhaustive bidirectional search algorithm for block matching, called Full Bidirectional Search (FBDS). It uses zero motion prejudgement inspired from Adaptive Rood Pattern Search (ARPS). The tool performs bidirectional motion compensation (BDMC) on the GPU using OpenCL. Testing was performed on three high-definition videos recorded by us,fitting the use scenario of remote surveillance. Results show increased quality compared to simple frame averaging, although with occasional block artefacts. The tool is capable of up-converting the recorded videos by a factor of three in a timespan less than 50 ms. Hence, it is viable for use in a high-definition remote surveillance system.