# Point Cloud Clustering System with DBSCAN Algorithm for Low-Resolution LiDAR

## Abstract

LiDAR point cloud clustering is a crucial part of object detection and recognition. However, clustering enormous point cloud of LiDAR assigns a large processing load to an on-board device in a vehicle. In this paper, we propose point cloud clustering system with a density-based spatial clustering of applications with noise (DBSCAN) algorithm for low-resolution LiDAR, offloading clustering tasks and shortening the processing time. In order to verify the feasibility of the system, we implemented the point cloud clustering accelerator on a field programmable gate array (FPGA). The system demonstrated 39.5 times enhancement in the processing speed.

## Authors

Sangho Lee *Dept. of Electronic Engineering, Seoul National University of Science and Technology, Seoul, Republic of Korea*

Seongmo An *Dept. of Electronic Engineering, Seoul National University of Science and Technology, Seoul, Republic of Korea*

Raehyeong Kim *Dept. of Electronic Engineering, Seoul National University of Science and Technology, Seoul, Republic of Korea*

Jongwon Oh *Dept. of Electronic Engineering, Seoul National University of Science and Technology, Seoul, Republic of Korea*

Seung Eun Lee *Dept. of Electronic Engineering, Seoul National University of Science and Technology, Seoul, Republic of Korea*

## Publication Information

**Journal:** 2024 IEEE International Conference on Consumer Electronics (ICCE) **Year:** 2024 **Pages:** 1-2 **DOI:** [10.1109/ICCE59016.2024.10444271](https://doi.org/10.1109/ICCE59016.2024.10444271) **Article Number:** 10444271 **ISSN:** Electronic ISSN: 2158-4001, Print on Demand(PoD) ISSN: 2158-3994

## Metrics

**Paper Citations:** 1 **Total Downloads:** 328

## Funding

- Korea Institute for Advancement of Technology

---

## Keywords

**IEEE Keywords:** Point cloud compression, Laser radar, Clustering algorithms, Object detection, Logic gates, Task analysis, Field programmable gate arrays

**Index Terms:** Point Cloud, Point Cloud Clustering, Low-resolution LiDAR, Processing Speed, Object Detection, Object Recognition, Density-based Clustering, LiDAR Point Clouds, Computation Time, Data Storage, Data Transmission, System Architecture, Control Mode, Computational Load, Image Sensor, Edge Computing, Coordinate Data, Hardware Accelerators, LiDAR Sensor, Pixel Index, Datapath

**Author Keywords:** Point Cloud Clustering system, DBSCAN, Li-DAR, FPGA

undefined
## SECTION I. Introduction

Recently, object detection and object recognition technology through LiDAR sensors or image sensors have been studied for autonomous driving. Clustering is crucial for object detection and recognition, and, particularly, LiDAR point cloud clustering identifies object positions and sizes, and even generates bounding boxes for two-stage detection [^1].

One of the clustering algorithms which clusters the LiDAR point cloud is the density-based spatial clustering of applications (DBSCAN) algorithm. The DBSCAN algorithm employs density-based clustering unlike the K-Means clustering and the Hierarchical clustering utilizing distances between clusters. Clusters are generated as long as a minimum number of points are present within a specified radius based on a randomly selected point. Additionally, when the criterion point of a cluster is encompassed by another cluster, the two clusters are merged into one. If a point is excluded from all groups, the point is handled as a noise. These characteristics of DBSCAN enable the LiDAR point cloud to be clustered [^2].

Clustering a substantial LiDAR point cloud involves essential iterative computations that expend significant time, hindering real-time object recognition in autonomous driving. This challenge is especially notable in on-board devices where diverse sensors are integrated. Dealing with massive point clouds creates sizeable computational loads, and even utilizing cloud computing is constrained by data size, leading to communication delays affecting sensor responsiveness [^3].

Edge computing technology involves computing data in close proximity to the sensors. In the context of vehicles, this approach reduces the computational load on on-board devices by designing systems that compute data from LiDAR sensors or image sensors. Furthermore, the existence of accelerators that operate based on data from LiDAR or images within the system further decreases computation time. In previous research, there are instances where the DBSCAN clustering method was applied to field programmable gate arrays(FPGA), successfully implementing the algorithm and achieving time saving [^4].

![Figure 1](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/10444098/10444131/10444271/lee1-1570945857-large.gif)

*Fig. 1. The block diagram of point cloud clustering system*

In this paper, we propose a point cloud clustering system with the DBSCAN algorithm for low-resolution LiDAR. This system transforms data near the sensors through edge computing, thereby alleviating the computational burden on the on-board device and reducing processing time by transmitting only the results and incorporating a LiDAR point cloud clustering hardware accelerator. In order to mitigate the computational load stemming from the utilization of extensive data, we employed low-resolution LiDAR for clustering and executed pre-processing steps such as integer quantization.

## SECTION II. System Architecture

Fig. 1 shows an overall architecture of the proposed system. The system consists of a microcontroller unit (MCU), an SRAM, and the LiDAR point cloud clustering accelerator. And the accelerator contains an (serial peripheral interface) SPI controller, an SRAM controller, a data path controller, and a grouping core.

The MCU quantizes x, y, and z coordinate data from the LiDAR sensor’s point cloud, sorts pixel indexes by grouping priority and sends data via SPI to the accelerator. Also, the MCU identifies group positions and sizes based on accelerator-transmitted group data. The coordinate data is quantized to 8-bit integers for efficient computing before transmission, and pixel indexes are aligned with normal vectors and distances from pixel coordinate data in the MCU for clustering accuracy. The accelerator addresses MCU-transmitted data for clustering and transmits the resultant group data.

![Figure 2](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/10444098/10444131/10444271/lee2-1570945857-large.gif)

*Fig. 2. Experimental environment*

In the grouping core, clustering acceleration is mainly advanced, while other modules handle data transmission or storage. The SPI controller module receives and transmits coordinate data and a sorted pixel list between the MCU and the accelerator. The SRAM controller module manages data storage from the data path controller, including clustering results. The data path controller module facilitates data transmission among the SPI and SRAM controller, and Grouping core modules, handling data reception, storage, and transmission.

Clustering based on the DBSCAN method is excuted in the grouping core module. This module utilizes indexes of the LiDAR’s pixels to adapt the DBSCAN method. Starting with a reference point from the sorted pixel index list, nearby pixels are sequentially compared with a reference point to evaluate point density. By placing the register array in the grouping core, it is possible to confirm whether the pixel was already included in a group, reducing the clustering performance time. Each time a group is created, the group number and indexes of pixels are saved in SRAM along with the index of the pixel.

## SECTION III. Experiment

We established a system comprising a point cloud clustering accelerator deployed on both Raspberry Pi and FPGA for system validation as shown in Fig. 2 (a). Groups generated by the system are shown on Fig. 2 (b). The Pixell LiDAR dataset from Pixset was utilized in the experiment, encompassing a total of 189 frames [^5]. To verify the feasibility of the system, we compare the processing time of a system designed with python framework on CPU and a system with HW accelerator, what we propose. When clustering was executed for each system, the time consumption was recorded at an average of 87.6 msec per frame and 2.2 msec per frame, respectively shown on Fig. 3. As a result, the FPGA-based system demonstrated a speed improvement of 39.5x in clustering computation time compared to the SW-based implementation. Hence, through comparative analysis of the experimental results, the designed hardware accelerator within the proposed system effectively demonstrated successful time reduction.

![Figure 3](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/10444098/10444131/10444271/lee3-1570945857-large.gif)

*Fig. 3. Experimental results*

## SECTION IV. Conclusion

In this paper, we propose the point cloud clustering system for low-resolution LiDAR. The proposed system carries out efficient computation for a short period of time through accelerators. In addition, we reduce the computational burden of the device by delivering only the results of clustering to an onboard device. By comparing and verifying the clustering time of the system implemented with SW and the system which utilizes FPGA, we demonstrated that the hardware accelerator in the proposed system effectively reduces the computation time.

## References

[^1]: F. Gao, C. Li and B. Zhang, “A Dynamic Clustering Algorithm for Lidar Obstacle Detection of Autonomous Driving System,” in IEEE Sensors Journal, vol. 21, no. 22, pp. 25922–25930, 2021. [IEEE](https://ieeexplore.ieee.org/document/9562508) [Google Scholar](https://scholar.google.com/scholar?as_q=A+Dynamic+Clustering+Algorithm+for+Lidar+Obstacle+Detection+of+Autonomous+Driving+System&as_occt=title&hl=en&as_sdt=0%2C31)

[^2]: X. Zhang and X. Huang, “Real-Time Fast Channel Clustering for LiDAR Point Cloud,” in IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 69, no. 10, pp. 4103–4107, 2022. [IEEE](https://ieeexplore.ieee.org/document/9803249) [Google Scholar](https://scholar.google.com/scholar?as_q=Real-Time+Fast+Channel+Clustering+for+LiDAR+Point+Cloud&as_occt=title&hl=en&as_sdt=0%2C31)

[^3]: X. Shen, Z. Chang, and S. Niu, “Mobile Edge Computing Task Offloading Strategy Based on Parking Cooperation in the Internet of Vehicles,” Sensors, vol. 22, no. 13, p. 4959, 2022. [DOI](https://doi.org/10.3390/s22134959) [Google Scholar](https://scholar.google.com/scholar?as_q=Mobile+Edge+Computing+Task+Offloading+Strategy+Based+on+Parking+Cooperation+in+the+Internet+of+Vehicles&as_occt=title&hl=en&as_sdt=0%2C31)

[^4]: Yizhao Gao, Song Wang, and Hayden Kwok-Hay So, “REMOT: A Hardware-Software Architecture for Attention-Guided Multi-Object Tracking with Dynamic Vision Sensors on FPGAs,” 2022 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA ’22), pp. 158–168, 2022. [DOI](https://doi.org/10.1145/3490422.3502365) [Google Scholar](https://scholar.google.com/scholar?as_q=REMOT%3A+A+Hardware-Software+Architecture+for+Attention-Guided+Multi-Object+Tracking+with+Dynamic+Vision+Sensors+on+FPGAs&as_occt=title&hl=en&as_sdt=0%2C31)

[^5]: J. Déziel et al., “PixSet: An Opportunity for 3D Computer Vision to Go Beyond Point Clouds With a Full-Waveform LiDAR Dataset,” 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pp. 2987–2993, 2021. [IEEE](https://ieeexplore.ieee.org/document/9565047) [Google Scholar](https://scholar.google.com/scholar?as_q=PixSet%3A+An+Opportunity+for+3D+Computer+Vision+to+Go+Beyond+Point+Clouds+With+a+Full-Waveform+LiDAR+Dataset&as_occt=title&hl=en&as_sdt=0%2C31)

### Additional References