Signal Clustering With Class-Independent Segmentation

Abstract

Radar signals have been dramatically increasing in complexity, limiting the source separation ability of traditional approaches. In this paper we propose a Deep Learning-based clustering method, which encodes concurrent signals into images, and, for the first time, tackles clustering with image segmentation. Novel loss functions are introduced to optimize a Neural Network to separate the input pulses into pure and non-fragmented clusters. Outperforming a variety of baselines, the proposed approach is capable of clustering inputs directly with a Neural Network, in an end-to-end fashion.

Machine Learning and Deep Learning Techniques for Colocated MIMO Radars: A Tutorial Overview

“…It is important to note that the application of the above mentioned DL techniques to object detection and localization in radar images is still at an early stage. Despite this, specific DL methods inspired by FCNs and U-Net have been already implemented for detecting and estimating the position of different targets (like cars and other automotive targets) on the basis of range-Doppler-azimuth radar maps [148]- [150]. Moreover, the use of semantic segmentation in the radar field has been already investigated for the classification and localization of 3D point clouds of automotive targets, like cars, tractors and pedestrians; various results referring to automotive MIMO radars that operate at 77 GHz can be found in refs.…”

Panoster: End-to-end Panoptic Segmentation of LiDAR Point Clouds

“…With the proposed Panoster, we simplify by removing this extra step, and incorporating the clustering in the network itself. We achieve this with an entirely learning-based approach, adapted from the signal processing domain [11], to output instance IDs directly, for each point given in input (Section III-C). To obtain panoptic predictions, it is then sufficient to filter out these IDs for all stuff points (Section III-D), as shown in Fig.…”

“…Our instance branch addresses its task as class-agnostic semantic segmentation, making a prediction for each point. In [11] the authors proposed a learning-based clustering method to separate radar signals by source. We adapt this approach to suit the panoptic task: cluster elements are 3D points, and each ground truth cluster represents an object instance.…”

“…By considering the softmax probabilities as cluster assignment probabilities, confusion matrices become soft confusion matrices S, allowing for the construction of loss functions. We refer the reader to [11] for further details.…”

“…We name our method Panoster, standing for panoptic clustering. Panoster brings panoptic capabilities to semantic segmentation networks by adding an instance branch trained to resolve objects with a learning-based clustering method [11]. Unlike existing proposal-free approaches [6,10], Panoster, thanks to its clustering solution integrated in the model, does not need further groupings to form objects, as it outputs instance IDs straight away from the network.…”