A Perception Engine in Automotive FMCW Radar
Kaluri V. Rangarao, Atul Negi · 2023
This paper introduces a new lean perception engine focusing on small-footprint software, which is essential for automotive and mobile device applications. The method involves quantifying a scene in a vertical plane over a short period by estimating the number of objects and associated parameters for each object. Real-time raw data is collected after each chirp using a single board FMCW radar sensor like AWR1642 under various scenarios. This raw data is then converted into a cloud of range (r) and Direction Of Arrival (DOA) (φ) by using gold-Fourier-Kaluri (gFK) and (κMUSIC) respectively. Object centroid $\bar{x}, \bar{y}$ and its statistics $\sigma_{x}$ and $\sigma_{y}$ are estimated from the cloud data. Lean partitioning and clustering are performed on this cloud data to obtain the ith object feature vector $s_{i} = [\bar{r}, \bar{\varphi} \delta r \delta\varphi]$. The composite method provides a processed scene matrix $\mathbf{S}(t)$, with $i_{t h}$ row $s_{i}$. The method details and sample results are presented here. This $\mathbf{S}(t)$ can be used for Machine Learning (ML) algorithms for classifying objects and tracking algorithms.