RADDet: Range-Azimuth-Doppler based Radar Object Detection for Dynamic Road Users (2105.00363v1)

Abstract: Object detection using automotive radars has not been explored with deep learning models in comparison to the camera based approaches. This can be attributed to the lack of public radar datasets. In this paper, we collect a novel radar dataset that contains radar data in the form of Range-Azimuth-Doppler tensors along with the bounding boxes on the tensor for dynamic road users, category labels, and 2D bounding boxes on the Cartesian Bird-Eye-View range map. To build the dataset, we propose an instance-wise auto-annotation method. Furthermore, a novel Range-Azimuth-Doppler based multi-class object detection deep learning model is proposed. The algorithm is a one-stage anchor-based detector that generates both 3D bounding boxes and 2D bounding boxes on Range-Azimuth-Doppler and Cartesian domains, respectively. Our proposed algorithm achieves 56.3% AP with IOU of 0.3 on 3D bounding box predictions, and 51.6% with IOU of 0.5 on 2D bounding box prediction. Our dataset and the code can be found at https://github.com/ZhangAoCanada/RADDet.git.

• The paper introduces a deep learning model and novel RAD tensor dataset that enable robust radar-based detection for dynamic road users.
• The proposed one-stage anchor-based detector achieves 56.3% AP for 3D bounding boxes and 51.6% for 2D boxes.
• It leverages a customized RadarResNet and automatic stereo vision annotations to enhance performance in challenging conditions.

Overview of RADDet: Range-Azimuth-Doppler Based Radar Object Detection for Dynamic Road Users

The paper "RADDet: Range-Azimuth-Doppler Based Radar Object Detection for Dynamic Road Users" ventures into the relatively underexplored domain of automotive radar-based object detection using deep learning models. The authors highlight the robustness of radar sensors in contrast to cameras and LiDARs, citing resistance to adverse weather conditions and affordability as significant advantages. The paper addresses the pressing issue of the limited availability of public radar datasets and proposes a comprehensive dataset along with a novel deep learning model for object detection utilizing Range-Azimuth-Doppler (RAD) tensor data.

Contributions and Methodology

The authors make several notable contributions. Firstly, they introduce a novel radar dataset represented as RAD tensors, paired with bounding box annotations and category labels for dynamic road users. These annotations are augmented by an automatic annotation method that enriches the data quality by leveraging stereo vision for category labeling. The automatic annotation increases the detection rate for dynamic objects by analytically connecting RAD patterns using enhanced detection techniques such as DBSCAN.

Secondly, the paper proposes a novel multi-class object detection model using radar data as a primary input. The model, designed as a one-stage anchor-based detector, predicts 3D bounding boxes in the RAD domain and 2D bounding boxes in the Cartesian domain. The architecture is built upon a customized RadarResNet backbone, chosen through systematic exploration and evaluated against state-of-the-art image-based object detection algorithms. The model achieves an average precision (AP) of 56.3% at an IoU threshold of 0.3 for 3D boxes and 51.6% at IoU 0.5 for 2D boxes, demonstrating competitive performance.

Key Numerical Results

• The model achieves an AP of 56.3% for 3D bounding box predictions at an IoU threshold of 0.3, and 51.6% for 2D bounding box predictions at an IoU threshold of 0.5.
• Extensive backbone explorations identified the RadarResNet with repeated residual blocks to provide the most optimal performance, outperforming alternatives including VGG-based architectures.

Implications and Future Directions

This research lays a groundwork for further exploration in the radar-based object detection domain. RADDet demonstrates that radar sensors, when paired with deep learning models, are viable for object detection tasks, challenging the dominant use of image-based methodologies. The findings suggest that enhanced radar object detection could complement existing camera and LiDAR systems in autonomous driving applications, especially in challenging environmental scenarios.

Future research can focus on expanding and diversifying the dataset to improve model accuracy across all categories of road users. Additionally, deploying such models on embedded platforms might require specialized optimization to maintain performance without hardware constraints hindering scalability.

Conclusively, RADDet not only fosters advancements in radar-based object detection but also encourages the development of algorithms that harness the full potential of radar technology. Such advancements could play a crucial role in the evolution towards more robust and inclusive autonomous systems.