Evaluation of Different Decision Fusion Mechanisms for Robust Moving Object Detection

A fundamental processing step in a wide variety of video processing pipelines is detection of moving objects. Moving object detection is a challenging task due to various environmental conditions such as illumination changes, shadows, glare, background clutter; foreground complexities such as occlusion, camouflage, complex motion behavior of the foreground objects; and imaging conditions such as low resolution and/or frame rate, camera jitter etc. While many moving object detection methods have been proposed, individual approaches often fail to address all the challenges efficiently. In this paper, we propose and evaluate different decision fusion mechanisms for robust detection of moving objects. The proposed hybrid system relies on motion, change, appearance cues and utilizes a mix of classical unsupervised computer vision and supervised deep learning approaches. Motion and change cues are computed through a tensor based motion estimation and a multi-modal background subtraction modules. Appearance cue is estimated using an unsupervised deep semantic segmentation network. These complementary visual cues are fused to achieve robust performance under challenging real-world conditions. The proposed multi-cue decision fusion pipelines were tested and evaluated on the CDnet-2014 change detection dataset. Our novel, unsupervised, expert-guided, semantic rule-based fusion approach with global channel filtering (SR-Fusion+GCF) is shown to outperform the other decision fusion strategies with an F-Measure of 69%, a promising 12% increase in F-measure compared to the best single cue approach.