Graph-Based Optimization and Statistical Inference for Scalable Multi-Camera Video Analytics

Abstract

Multi-camera video analytics involves robustly associating video identities across cameras and keeping computational cost down as more cameras and observations are added. For scalable multi-camera video analytics,this paper presents an Uncertainty-Aware Adaptive Graph Optimization and Statistical Inference (UA GO-SI) framework. Each observation of an object is shown as a node in a dynamic spatio-temporal graph, and the relationships between objects in the two cameras are modeled through probabilistic edges. The four factors, appearance similarity, temporal consistency, motion consistency and learnt camera-transition probability, are fused together by adaptive modality weights. To measure uncertainty at the edge level, the association probability is then mapped to a binary entropy and normalized. This retention of graph edges then becomes an optimization problem of budget considering all of the above. Association probabilities, uncertainty estimates, selection of edges and adaptive weights are periodically changed by an alternating optimization procedure until convergence is achieved. In addition, there is also a theoretical sparsification condition that ties removing low information edges to a bounded drop-off in an objective. The evaluation design is geared towards the analysis of tracking accuracy, identity consistency, uncertainty calibration, graph sparsity, computational cost, scalability, robustness, camera failure, temporal stability, and convergence. The goal is to produce a compromise solution that keeps the high information/low uncertainty associations intact and prunes out redundant or ambiguous relations. The formulation that you see is the unified one for scalable identity inference on a computational level, with no fabricated experimental measurements.