Date of Award


Publication Type

Doctoral Thesis

Degree Name



Electrical and Computer Engineering

First Advisor

Wu, Qingming


Background subraction, Deep convolutional neural network, Deep learning, Video foreground




These days, detection of Visual Attention Regions (VAR), such as moving objects has become an integral part of many Computer Vision applications, viz. pattern recognition, object detection and classification, video surveillance, autonomous driving, human-machine interaction (HMI), and so forth. The moving object identification using bounding boxes has matured to the level of localizing the objects along their rigid borders and the process is called foreground localization (FGL). Over the decades, many image segmentation methodologies have been well studied, devised, and extended to suit the video FGL. Despite that, still, the problem of video foreground (FG) segmentation remains an intriguing task yet appealing due to its ill-posed nature and myriad of applications. Maintaining spatial and temporal coherence, particularly at object boundaries, persists challenging, and computationally burdensome. It even gets harder when the background possesses dynamic nature, like swaying tree branches or shimmering water body, and illumination variations, shadows cast by the moving objects, or when the video sequences have jittery frames caused by vibrating or unstable camera mounts on a surveillance post or moving robot. At the same time, in the analysis of traffic flow or human activity, the performance of an intelligent system substantially depends on its robustness of localizing the VAR, i.e., the FG. To this end, the natural question arises as what is the best way to deal with these challenges? Thus, the goal of this thesis is to investigate plausible real-time performant implementations from traditional approaches to modern-day deep learning (DL) models for FGL that can be applicable to many video content-aware applications (VCAA). It focuses mainly on improving existing methodologies through harnessing multimodal spatial and temporal cues for a delineated FGL. The first part of the dissertation is dedicated for enhancing conventional sample-based and Gaussian mixture model (GMM)-based video FGL using probability mass function (PMF), temporal median filtering, and fusing CIEDE2000 color similarity, color distortion, and illumination measures, and picking an appropriate adaptive threshold to extract the FG pixels. The subjective and objective evaluations are done to show the improvements over a number of similar conventional methods. The second part of the thesis focuses on exploiting and improving deep convolutional neural networks (DCNN) for the problem as mentioned earlier. Consequently, three models akin to encoder-decoder (EnDec) network are implemented with various innovative strategies to improve the quality of the FG segmentation. The strategies are not limited to double encoding - slow decoding feature learning, multi-view receptive field feature fusion, and incorporating spatiotemporal cues through long-shortterm memory (LSTM) units both in the subsampling and upsampling subnetworks. Experimental studies are carried out thoroughly on all conditions from baselines to challenging video sequences to prove the effectiveness of the proposed DCNNs. The analysis demonstrates that the architectural efficiency over other methods while quantitative and qualitative experiments show the competitive performance of the proposed models compared to the state-of-the-art.