Optimisation of video detection algorithms for use in advertisement detection applications
The focus of this thesis is on the optimisation of the video detection process to detect advertisements in streaming digital media. While the process of video detection has been around for many years, it is very limited in its scope of application. These limitations pertain not only to the types of video, but in particular the execution time thereof. For a video detection technique to be effectively utilised in an advertisement detection environment, it must be able to perform real-time analysis. A proposed method was found in literature which addressed some of the core functionality required for such an application. This method was however still extremely limited with regard to its scope of application and devoid of scientific justification for the parameters used therein. The work presented in this thesis aim to address these limitations by not only expanding the scope of operation of the detection algorithm, but to provide scientific justification for the techniques and parameters utilised therein. One of these core components is the video segmentation algorithm, realised by employing the Jensen-Shannon divergence. While the Jensen-Shannon divergence is commonly seen as an information metric, it poses an uncanny ability to help detect video shot boundaries. By analysing the Jensen-Shannon divergence in this context, a profound insight into the technique and its associated parameters was derived. In doing so, a wider variety of videos can be detected with better accuracy and precision. Since the video segmentation aspect of the investigation provided a means to increase the speed by reliably reducing the data to be processed, the subsequent core module tasked with identifying the video was evaluated. The identification of an unknown video is done by extracting and comparing a unique, yet robust, digital video fingerprint against a known database. Due to the infinite variance in possible advertisements, a unique video fingerprinting algorithm was employed, consisting of unique hash descriptors derived from prominent keypoints found within each video. While each of these core modules have been individually optimised, the main contribution of this thesis is the integration of these modules to create a video detection ecosystem incorporating the unique underlying dependencies between these modules. Lastly, the optimisation of the integrated video detection ecosystem, provides a means of detecting a multitude of different videos, while adhering to the real-time processing requirements with scientific justification for the techniques and parameters employed to accomplish the task.
- Engineering