Radar-vision fusion for object classification

We propose an object classification system that incorporates information from a video camera and an automotive radar. The system implements three processes. The first process is attention selection, in which the radar guides a selection of a small number of candidate images for analysis by the camera and our learning method. In the second process, normalized attention windows are processed by orientation-selective feature detectors, generating a sparse representation for each window. In the final process, a multilayer in-place learning network is used to distinguish sparse representations of different objects. Though it is more flexible in terms of variety of classification tasks, the system currently demonstrates its high accuracy in comparison with others on real-world data of a two-class recognition problem.