Access Type

Open Access Dissertation

Date of Award

January 2022

Degree Type


Degree Name



Electrical and Computer Engineering

First Advisor

Nabil Sarhan


This dissertation considers computer vision (CV) systems in which a central monitoring station receives and analyzes the video streams captured and delivered wirelessly by multiple cameras. It addresses how the bandwidth can be allocated to various cameras by presenting a cross-layer solution that optimizes the overall detection or recognition accuracy. The dissertation presents and develops a real CV system and subsequently provides a detailed experimental analysis of cross-layer optimization. Other unique features of the developed solution include employing the popular HTTP streaming approach, utilizing homogeneous cameras as well as heterogeneous ones with varying capabilities and limitations, and including a new algorithm for estimating the effective medium airtime. The results show that the proposed solution significantly improves the CV accuracy.

Additionally, the dissertation features an improved neural network system for object detection. The proposed system considers inherent video characteristics and employs different motion detection and clustering algorithms to focus on the areas of importance in consecutive frames, allowing the system to dynamically and efficiently distribute the detection task among multiple deployments of object detection neural networks. Our experimental results indicate that our proposed method can enhance the mAP (mean average precision), execution time, and required data transmissions to object detection networks.

Finally, as recognizing an activity provides significant automation prospects in CV systems, the dissertation presents an efficient activity-detection recurrent neural network that utilizes fast pose/limbs estimation approaches. By combining object detection with pose estimation, the domain of activity detection is shifted from a volume of RGB (Red, Green, and Blue) pixel values to a time-series of relatively small one-dimensional arrays, thereby allowing the activity detection system to take advantage of highly capable neural networks that have been trained on large GPU clusters for thousands of hours. Consequently, capable activity detection systems with considerably fewer training sets and processing hours can be built.