The sparse nature of the ranging and spatial angle
parameter space has been exploited by many radar parameter
estimation algorithms in literature. We note that real world
reflections are not sporadically sparse in the parameter space and
typically exhibit smooth variation effects with non-zero entries
occurring in clusters. In this paper, we explicitly model this
additional structural information into our estimation algorithm
and propose a non-convex regularization of the linear observation
model corresponding to the sparse dictionary representation of
the radar image. The proposed non-convex regularization encourages
block-sparse solutions and incorporates side information
from a secondary camera sensor. In addition, the proposed
framework also allows us to use structural information learnt
using observations over time. Our experiments using a 77 GHz
FMCW radar only and radar-camera setup show improved
performance over state of the art algorithms in terms of target
detection and resolution.