Training-free moving object detection system based on hierarchical color-guided motion segmentation

We present a moving object detection system for surveillance based on Hierarchical Color-guided Motion segmentation (HiCoMo). The HiCoMo system does not require training and consists of two main stages: (1) hierarchical color-guided motion segmentation, and (2) motion-based verification. The first stage is a hierarchical segmentation framework, where at each level a balance is made between static and temporal features. So that groups of pixels develop into semantic object segments. In the second stage, these object segments are further analyzed in terms of motion saliency and consistency, in order to finalize the object detection results. Our proposed system is tested on real-life surveillance videos containing various scenarios. The detection results outperform a state-of-the-art training-free moving object detection algorithm in recall (90.2% compared to 81.6%) while having a competitively promising precision (96.5% compared to 97.4%). The system has a generic nature and real-time implementation potential, which makes it applicable to various applications of computer vision.