# 3D Object Detection Method Based on YOLO and K-Means for Image and Point Clouds

**Authors:** Xuanyu Yin, Yoko Sasaki, Weimin Wang, Kentaro Shimizu  
**arXiv:** [2005.02132](https://arxiv.org/abs/2005.02132)  
**Date:** 2020-04-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2005.02132v1/

## Key points

- Lidar-based 3D object detection and classification tasks in autonomous driving are important
- Real-time detection in 3D point clouds requires strong algorithmic support
- The paper proposes a novel method that combines point cloud and image data for 3D object detection
- The method consists of lidar-camera calibration, YOLO-based detection and PointCloud extraction, and K-means based point cloud segmentation
- Camera is used for real-time 2D object detection using YOLO, with bounding box information transferred to conduct 3D object detection on lidar point cloud data
- High-speed 3D object recognition functionality achieved on GPU by comparing 2D coordinate with object bounding box
- K-means clustering improves accuracy and precision of the detection method in point cloud data
- Proposed method offers faster speed compared to PointNet
- Comprehensive approach to lidar-based 3D object detection by integrating image and point cloud data
- Promising results in terms of accuracy, precision, and speed suitable for applications in autonomous driving systems

## Summary

The paper titled "3D Object Detection Method Based on YOLO and K-Means for Image and Point Clouds" discusses the importance of lidar-based 3D object detection and classification tasks in autonomous driving. The authors highlight that while lidar sensors can provide 3D point cloud data reconstruction of the surrounding environment, real-time detection in 3D point clouds still requires strong algorithmic support. To address this challenge, the paper proposes a novel 3D object detection method that combines point cloud and image data. The method consists of three main parts: (1) Lidar-camera calibration and undistorted image transformation, (2) YOLO-based detection and PointCloud extraction, and (3) K-means based point cloud segmentation and detection experiment test and evaluation in depth image. In their research, the authors utilize a camera to capture images for real-time 2D object detection using YOLO. They then transfer the bounding box information to a node responsible for conducting 3D object detection on point cloud data obtained from the lidar sensor. By comparing whether the 2D coordinate transferred from the 3D point lies within the object bounding box or not, they achieve high-speed 3D object recognition functionality on a GPU. The authors also highlight that k-means clustering improves the accuracy and precision of their detection method in point cloud data. Additionally, they emphasize that their proposed method offers faster speed compared to PointNet. Overall, this paper presents a comprehensive approach to lidar-based 3D object detection by integrating image and point cloud data. The proposed method shows promising results in terms of accuracy, precision, and speed which makes it suitable for applications in autonomous driving systems.