The field of computer vision has long strived to extract understanding from images and videos sequences. The recent flood of video data along with massive increments in computing power have provided the perfect environment to generate advanced research to extract intelligence from video data. Video data is ubiquitous, occurring in numerous everyday activities such as surveillance, traffic, movies, sports, etc. This massive amount of video needs to be analyzed and processed efficiently to extract semantic features towards video understanding. Such capabilities could benefit surveillance, video analytics and visually challenged people. While watching a long video, humans have the uncanny ability to bypass unnecessary information and concentrate on the important events. These key events can be used as a higher-level description or summary of a long video. Inspired by the human visual cortex, this research affords such abilities in computers using neural networks. Useful or interesting events are first extracted from a video and then deep learning methodologies are used to extract natural language summaries for each video sequence. Previous approaches of video description either have been domain specific or use a template based approach to fill detected objects such as verbs or actions to constitute a grammatically correct sentence. This work involves exploiting temporal contextual information for sentence generation while working on wide domain datasets. Current state-of- the-art video description methodologies are well suited for small video clips whereas this research can also be applied to long sequences of video.
This work proposes methods to generate visual summaries of long videos, and in addition proposes techniques to annotate and generate textual summaries of the videos using recurrent networks. End to end video summarization immensely depends on abstractive summarization of video descriptions. State-of- the-art neural language & attention joint models have been used to generate textual summaries. Interesting segments of long video are extracted based on image quality as well as cinematographic and consumer preference. This novel approach will be a stepping stone for a variety of innovative applications such as video retrieval, automatic summarization for visually impaired persons, automatic movie review generation, video question and answering systems.
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Kulhare, Sourabh, "Deep Learning for Semantic Video Understanding" (2017). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus