Towards Optimized IoT-based Context-aware Video Content Analysis Framework

Despite the success of convolutional neural networks (CNNs) in the area of spatial analysis, and recurrent neural networks (RNNs) on sequence modeling and interpretation tasks, video analysis has only seen limited interest and progress. This is partially due to focusing on the natural humanlike translation from video space to natural language space to the detriment of informativeness. This paper is proposing an automated context-aware video analysis framework that is directed by the constrains of its application. This framework encorporates an encoder-decoder neural network trained on a closed-domain video-to-text dataset. The network architecture and the standardized language model present in the dataset are optimized for speed, to allow the system to be applied on IoT devices, and for informativeness, to extract information easily from the model output to the following stages of the anlaysis. The proposed framework provides a practical method to integrate the power of CNN and RNN combination in a directed way to extract the most from video content. A classroom monitoring system is discussed as an example of the capabilities and limitations of the proposed framework using NVIDIA's Jetson nano board.