Radar Detection Using Deep Learning
(2022) In Master's Theses in Mathematical Sciences FMAM02 20221Mathematics (Faculty of Engineering)
- Abstract
- This thesis aims to reproduce and improve a paper about dynamic road user detection on 2D bird's-eye-view radar point cloud in the context of autonomous driving. We choose RadarScenes, a recent large public dataset, to train and test deep neural networks. We adopt the two best approaches, the image-based object detector with grid mappings approach and the semantic segmentation-based clustering approach. YOLO v3 and Pointnet++ are the deep networks for the two approaches, respectively. We implement an radar-based version of DBSCAN to extract instance clusters (objects). For both approaches, various preprocessing techniques are implemented, such as velocity skew function, upsampling and data augmentations, including rotation and flipping. We... (More)
- This thesis aims to reproduce and improve a paper about dynamic road user detection on 2D bird's-eye-view radar point cloud in the context of autonomous driving. We choose RadarScenes, a recent large public dataset, to train and test deep neural networks. We adopt the two best approaches, the image-based object detector with grid mappings approach and the semantic segmentation-based clustering approach. YOLO v3 and Pointnet++ are the deep networks for the two approaches, respectively. We implement an radar-based version of DBSCAN to extract instance clusters (objects). For both approaches, various preprocessing techniques are implemented, such as velocity skew function, upsampling and data augmentations, including rotation and flipping. We also adapt the evaluation metrics, IOU, mAP, and F1-score for point clusters so that both approaches' output can be comparable. The reproduction of both approaches achieved comparable performance as in the original paper, which indicates the image-based detector overwhelmed the semantic segmentation-based clustering approach. We also managed to improve the metrics by adapting clever variations in the DBSCAN pipeline. Besides, we implemented the ablation study for the YOLO approach and found horizontal flipping the point cloud as the optimal data augmentation operation. We implemented the ablation study for the PointNet/DBSCAN pipeline as well and found that randomly jittering the points considering the radial velocity of the radar reflections output the best model, and in under specific cases, it improved it. We also investigated the effect of time accumulation on APs of all the classes. We found that low AP of the pedestrian class is the performance bottleneck, and simply accumulating a longer period cannot significantly improve it. (Less)
- Popular Abstract
- Nowadays, when one interacts with a smartphone and the installed apps that recognize our faces or detect songs just by listening to a little extract of a piece, one is exposed to machine/deep learning models without notice. The same techniques have been exploited in the not-to-old advanced technology of self-driving cars. These techniques have gifted the vehicles with the ability to make decisions on detected road objects which helps us to have a safely and comfortably driving environment.
Traditionally, the machine-learning-based object detection field has been a camera-based domain.
This phenomenon occurs for object detection in cars as well. Thus, scientists and engineers have developed many object detector systems that use cameras... (More) - Nowadays, when one interacts with a smartphone and the installed apps that recognize our faces or detect songs just by listening to a little extract of a piece, one is exposed to machine/deep learning models without notice. The same techniques have been exploited in the not-to-old advanced technology of self-driving cars. These techniques have gifted the vehicles with the ability to make decisions on detected road objects which helps us to have a safely and comfortably driving environment.
Traditionally, the machine-learning-based object detection field has been a camera-based domain.
This phenomenon occurs for object detection in cars as well. Thus, scientists and engineers have developed many object detector systems that use cameras and have achieved outstanding performances in recent years. However, for radars, the story is different. Since radar is typically known for providing less rich information than a camera or Lidar sensors, the machine learning algorithms were likelier to perform poorly when detecting objects in space using radar data. Therefore, the
technology was not studied enough, and the radar was only used as a side component that helped other car sensors.
With the breakthrough in artificial intelligence, the self-driving automobile is no longer a dream in science fiction. One of the most crucial tasks of a self-driving car is to detect other road users so that the car can make decisions for the car to maneuver accordingly. This task requires the car to localize other road users and determine their types simultaneously.
In order to confront such a challenge, the car is equipped with different types of sensors, such as cameras, Lidars, and radars. Among three mainstream types of sensors, radars are the only ones that can measure not only the object distance but also the relative velocity. This is nontrivial!
The knowledge of whether the other road users are approaching us or not will lead to different car maneuver operations. However, the radar's output is a point cloud, which is overly sparse and even difficult for human eyes to see directly.
Moreover, the new AI algorithms and richer-featured automotive radar sensors could greatly improve the way radars are being used in automated driver tasks. This fact has advantages, including reducing prices in sensory systems for self-driving car technology and providing robustness to the systems mentioned above.
In this thesis, we work with radar point-cloud data for the purpose of object detection. To overcome the sparsity, we turn to deep learning algorithms, which consist of a main branch of AI. Deep learning is a powerful tool that is not based on a set of deterministic rules; instead, it can learn from big data.
We intend with this work to test the limits of radar with the fanciest and most promising deep learning architectures and generate reliable object detectors that could also be enlightenment for future research works in this field. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9098412
- author
- Xiong, Ziliang LU and Carrera, Leonardo
- supervisor
- organization
- course
- FMAM02 20221
- year
- 2022
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Deep Learning, Autonomous Driving, Radar, Point Cloud
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMA-3488-2022
- ISSN
- 1404-6342
- other publication id
- 2022:E65
- language
- English
- id
- 9098412
- date added to LUP
- 2022-09-19 16:56:26
- date last changed
- 2022-09-19 16:56:26
@misc{9098412,
abstract = {{This thesis aims to reproduce and improve a paper about dynamic road user detection on 2D bird's-eye-view radar point cloud in the context of autonomous driving. We choose RadarScenes, a recent large public dataset, to train and test deep neural networks. We adopt the two best approaches, the image-based object detector with grid mappings approach and the semantic segmentation-based clustering approach. YOLO v3 and Pointnet++ are the deep networks for the two approaches, respectively. We implement an radar-based version of DBSCAN to extract instance clusters (objects). For both approaches, various preprocessing techniques are implemented, such as velocity skew function, upsampling and data augmentations, including rotation and flipping. We also adapt the evaluation metrics, IOU, mAP, and F1-score for point clusters so that both approaches' output can be comparable. The reproduction of both approaches achieved comparable performance as in the original paper, which indicates the image-based detector overwhelmed the semantic segmentation-based clustering approach. We also managed to improve the metrics by adapting clever variations in the DBSCAN pipeline. Besides, we implemented the ablation study for the YOLO approach and found horizontal flipping the point cloud as the optimal data augmentation operation. We implemented the ablation study for the PointNet/DBSCAN pipeline as well and found that randomly jittering the points considering the radial velocity of the radar reflections output the best model, and in under specific cases, it improved it. We also investigated the effect of time accumulation on APs of all the classes. We found that low AP of the pedestrian class is the performance bottleneck, and simply accumulating a longer period cannot significantly improve it.}},
author = {{Xiong, Ziliang and Carrera, Leonardo}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{Radar Detection Using Deep Learning}},
year = {{2022}},
}