Adaptive Segmentation of Moving Objects Using Spatio-Temporal RAG Connectivity
This research presents a moving-object segmentation algorithm that extracts objects from video sequences without prior knowledge of their appearance, shape, or motion model. The method combines spatial segmentation with temporal evidence: a reference frame Iₜ is segmented into homogeneous regions, represented as a RAG = (V,E), and consecutive frames Iₜ₊ᵢ are compared iteratively to accumulate motion-related change information.
First Stage — Spatial Segmentation and RAG Construction
The reference frame Iₜ is first partitioned into semantic homogeneous regions. Each region becomes a node v ∈ V, and adjacency between neighboring regions becomes an edge e(u,v) ∈ E. This transforms the image into a structured graph whose topology preserves the object/background boundaries.
Second Stage — Multiple Temporal Comparisons
Instead of relying on only two frames, the algorithm compares Iₜ with a sequence of registered frames Iₜ₊₁,...,Iₜ₊ᴺ. Each comparison estimates a local change probability LCPₜ,ₜ₊ᵢ(e) on RAG boundary pixels. These repeated comparisons gradually distinguish true object edges from occluded or background edges.
Third Stage — Global Change Probability and Motion Evidence
The set of local probabilities is mapped into a global edge weight GCPᵢ(e), and then into node weights GCPᵢ(v). Object nodes tend to preserve high values across iterations, while occluded/background nodes become weaker. Thus, motion evidence is encoded directly on the graph.
Fourth Stage — Connectivity Analysis and Stabilization
A modified BFS-like Nodes Connectivity Analysis searches the weighted RAG over levels p = 100,...,0, contracts connected nodes, and extracts candidate moving-object sets. The temporal iterations stop when consecutive object masks become stable according to a boundary distance D(∂objᵢ, ∂objᵢ₋₁), meaning the segmentation has converged and locked onto the moving object.