Video Analytic Data Reduction Model for Edge Computing

The traditional approaches for video analytics use cloud computing for data processing. Employing deep learning in video analytics and transferring large volumes of data causes network congestion and high latency in centralized cloud computing systems. To minimize these problems, we propose and implement a Video Analytic Data Reduction Model (VADRM) that divides the video streaming jobs into smaller deep learning tasks for processing on edge computing. VADRM provides a distributed model for reducing processing requirements in each phase of the video analytics applications employing deep learning methods. In this dissertation, deep learning using the Convolutional Neural Network (CNN) is implemented for building a prototype of VADRM. We improve the object tracking phase by proposing a CNN-based object reidentification model in the VADRM prototype. The experimental results show that the proposed reidentification method increases the quality of object tracking compared to the existing methods. Simulation of video analytics is challenging because of the lack of real data generated by machine learning techniques in the new generations of these applications. In this dissertation, the iFogSim simulator is used with the real data of the VADRM prototype to evaluate the module deployment algorithms for the edge devices. The experimental results show that the edge network usage and latency are improved by 12% using our proposed module placement algorithm compared to the iFogSim Edge-ward algorithm. The data collected from the VADRM prototype are characterized to find the distribution models for developing the analytical and simulation models. Large artificial data are generated using these distribution models to conduct a resource management simulation comparing edge and cloud computing for deep learning-based video analytics applications.