Distributed compression and fusion of nonnegative sparse signals for multiple-view object recognition

Visual surveillance in complex urban environments requires an intelligent system to automatically track and identify multiple objects of interest in a network of distributed cameras. The ability to perform robust object recognition is critical to compensate adverse conditions and improve performance, such as multi-object association, visual occlusion, and data fusion with hybrid sensor modalities. In this paper, we propose an efficient distributed data compression and fusion scheme to encode and transmit SIFT-based visual histograms in a multi-hop network to perform accurate 3-D object recognition. The method harnesses an emerging theory of (distributed) compressive sensing to encode high-dimensional, nonnegative sparse signals via random projection, which is unsupervised and independent to the sensor modality. A multi-hop protocol then transmits the compressed visual data to a base-station computer, which preserves a constant bandwidth regardless of the number of active camera nodes in the network. Finally, the multiple-view object features are simultaneously recovered via ℓ¹-minimization as an efficient decoder. The efficacy of the algorithm is validated using up to four Berkeley CITRIC camera motes deployed in a realistic indoor environment. The substantial computation power on the CITRIC mote also enables fast compression of SIFT-type visual features extracted from object images.