3D 物體偵測之邊緣運算模組的設計與實現

Design and implementation of edge computing module for 3D object detection

何昀鴻(HO YUN HUNG)

指導教授 ： 李世安

淡江大學/工學院/電機工程學系人工智慧機器人碩士班/碩士(2023年)

https://doi.org/10.6846/TKU.2023.00031

摘要

本論文設計與實現一個三維光達(3D LiDAR)的邊緣運算模組，並且可以即時偵測周遭環境內的3D物體位置。在邊緣運算模組的部分，主要以Nvidia Jetson AGX Xavier作為運算平台，來即時處理三維光達的點雲資料。在3D物體偵測演算法部分主要分為2部分(1)點雲前處理與(2)物體偵測。由於三維光達每秒輸入的點雲資料約為3萬個點，若是要一次處理全部的點雲資料計算量過於龐大且耗時，所以在進行物體偵測前須先將資料進行濾除。本論文利用直通濾波器(passthrough filter)與體素網格濾波器(voxel grid filter)進行初步的點雲過濾，再利用隨機抽樣一致(RANSAC)演算法將點雲資料分為物體點雲與地面點雲且將地面點雲濾除。在物體偵測系統部分，以DBSCAN演算法作為點雲聚類之演算法，待聚類完成後為確認不同時間點下所偵測到的物體為同一物體，本文使用Gale-Shapley 演算法對點雲聚類匹配，最後透過計算物體的位移方向與位移量得以實現物體追蹤。在實驗結果的部分本論文之方法，在本論文之實驗場域中只需花費0.07759秒即可完成一次偵測。

關鍵字

三維光達 ； 邊緣運算 ； 體素網格濾波器 ； 物件偵測 ； 物件追蹤 ； 密度式分群演算法

並列摘要

In this thesis I had designed and implemented an edge computing module that can implement a 3D object detection system. In the part of the edge computing module, I use Nvidia Jetson AGX Xavier as the main execution computer, and use 3D LiDAR to obtain environmental object information. In terms of 3D object detection, it is mainly divided into two parts (1) point cloud pre-processing, (2) object detection system, since the point cloud data input by the 3D LiDAR is about 30,000 points per second, it would be too computationally intensive and time-consuming to process all the point cloud data at once, so the data must be filtered before object detection. This thesis uses passthrough filter and voxel grid filter for preliminary filtering, then use RANSAC algorithm to divide the point cloud data into object point cloud and ground and filter out the ground. In the part of the object detection system, the DBSCAN algorithm is used as the algorithm for point cloud clustering. After the clustering is completed, in order to confirm that the objects detected at different time points are the same object, the Gale-Shapley algorithm is used to match the point cloud clusters. Finally, the object tracking can be realized by calculating the displacement direction and displacement amount of the object. In the part of the results, our method only spend 0.07759 seconds to complete a detection in the experimental field of this paper.

並列關鍵字

3D LiDAR ； Edge Computing ； Voxel grid filter ； Object detection ； Object tracking ； DBSCAN

參考文獻

[1] 【雲端運算種類有哪些?】3分鐘詳解雲端服務類型與優勢, URL: https://www.easpnet.com/blog/cloud-services-101/ (2022/11/05 存取)

[2] 什麼是邊緣運算？, URL: https://www.intel.com.tw/content/www/tw/zh/edge-computing/what-is-edge-computing.html (2022/11/05 存取)

[3] AMR是什麼？一篇瞭解AMR自主移動機器人的趨勢與應用發展, URL: https://www.bnext.com.tw/article/68428/itri040602 (2022/11/05 存取)

[4] NVIDIA CLARA, URL: https://www.nvidia.com/zh-tw/clara/ (2022/11/08 存取)

[5] 台北市信義路公車專用道自駕巴士, URL: https://jobus.tw/turingdrive-xinyi/