Object Detection on Compressive Measurements using Correlation Filters and Sparse Representation

Compressive cameras acquire measurements of a scene using random projections instead of sampling at Nyquist rate. Several reconstruction algorithms have been proposed, taking advantage of previous knowledge about the scene. However, some inference tasks require to determine only certain information of the scene without incurring in the high computational reconstruction step. By reducing the computation load related to the reconstruction problem, this paper proposes a computationally efficient object detection approach based on correlation filters and sparse representation that operate over compressive measurements. We consider the problem of object detection in remote sensing scenes with multi-band images, where the pixels are expensive. The correlation filters are designed using explicit knowledge of the target appearance and the target shape to provide a way to recognize the objects from compressive measurements. Numerical experiments show the validity and efficiency of the proposed method in terms of peak-to-side lobe ratio using simulated data.