Hybrid Gaussian process visual target classification in wireless multimedia sensor networks

摘要

Visual target classification is one of the most important issues addressed in wireless multimedia sensor network (WMSN). This paper proposes a hybrid Gaussian process based classification method to implement binary visual classification (humanonhuman) in WMSN. Because the computation ability of sensor node in WMSN is strictly limited, target classification is achieved by Gaussian process classifier (GPC) with compacted features, where feature compaction is achieved by the combination of integer lifting wavelet transform and rough set theory. To improve the robustness and accuracy, this paper proposes to fuse the local decisions of multiple sensor nodes running GPCs with different kernel functions, where committee decision based hybrid decision fusion strategy is employed to combine the local decisions with dynamically adjusted weights. Experimental results verify that the proposed hybrid Gaussian process visual target classification method can effectively carry out target classification in WMSN. The committee decision based hybrid decision fusion strategy can reduce the impact of the uncertainty of classification and improve the overall performance.