Object detection for automotive radar point clouds – a comparison

Scheiner, Nicolas; Kraus, Florian; Appenrodt, Nils; Dickmann, Jürgen; Sick, Bernhard · 2021 · Crossref

DOI: 10.1186/s42467-021-00012-z

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of detecting and classifying moving road users using conventional automotive radar sensors, a critical component for automated driving systems. While radar offers superior robustness against adverse weather conditions compared to cameras and lidar, its data is inherently sparse, complicating object detection. The authors note that existing machine learning approaches for radar often lack comparative evaluation or rely on next-generation sensors unavailable in current vehicles. Consequently, this paper provides a thorough comparison of five real-time capable object detection architectures using a conventional 77 GHz radar system and a large, open real-world dataset. The experimental design utilizes a proprietary dataset comprising annotated bird’s-eye view point clouds with range, azimuth, amplitude, Doppler, and time information. The data covers five object classes (pedestrians, groups, bikes, cars, trucks) and background, captured by four front-mounted sensors. To ensure comparability, all methods process cropped frames of 500 ms in time and 100 m × 100 m in space. The study evaluates five distinct approaches: (1) a modular pipeline using two-stage clustering followed by a recurrent neural network (LSTM) classifier ensemble; (2) a semantic segmentation network (PointNet++) combined with class-sensitive clustering; (3) an image-based object detector (YOLOv3) adapted via grid mapping; (4) a point-cloud-based object detector (PointPillars); and (5) a hybrid combining semantic segmentation and clustering. Extensive preprocessing, including ego-motion compensation and coordinate transformation, is applied to all methods. The results indicate that the deep image-based detection network and the recurrent neural network ensemble outperform the other approaches. The image-based detector benefits from the dense representation created by grid mapping, while the LSTM ensemble effectively leverages temporal features through its sliding window approach. In contrast, the semantic segmentation and pure point-cloud methods showed lower performance, partly due to the sparsity of conventional radar data and the challenges in forming accurate clusters without prior classification or dense grid structures. The study also highlights that preprocessing steps, particularly point cloud filtering and clustering parameter optimization, significantly impact detection accuracy. The significance of this work lies in providing the first representative algorithm comparison for automotive radar object detection using conventional sensors. By evaluating diverse architectures on a common, large-scale dataset, the authors identify the strengths and limitations of current methods, offering clear guidance for future research. The findings suggest that hybrid approaches or those leveraging dense intermediate representations (like grid maps) are currently more effective for conventional radar systems. This evaluation helps bridge the gap between theoretical machine learning advancements and practical automotive applications, outlining specific directions for improving radar perception in automated driving.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms