Scholarlink

<p>Recently, mmWave Radar sensor has been used widely as perception sensor in autonomous vehicles due to its compact size and light weight. Typically, the mmWave Radar point cloud data generated are clustered using algorithms like Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, to associate the point cloud data to each of the obstacles present in the environment. However, the parameters for the DBSCAN algorithm are often fixed, and it does not take into account the varying environment encountered by the autonomous vehicles. Furthermore, the DBSCAN algorithm does not consider the radial velocity of each data point throughout the clustering process. All these problems can indeed lead to less precise clustering performance. Hence, this paper presented two statistical-based approaches to improve the precision of DBSCAN algorithm. The proposed K-Nearest Neighbors - Probability (KNN-Probability) method determines the <i>eps</i> parameter of DBSCAN algorithm dynamically for each point cloud data frame, while the proposed cluster refinement algorithm refines the DBSCAN clusters based on radial velocity data. By utilizing the publicly available RadarScenes dataset for performance evaluation, the proposed obstacle detection algorithm has achieved a more precise clustering result with a high mean average precision (mAP) of 0.8942. In terms of future research direction, the development of artificial intelligence (AI) based methods to optimize the selection of parameters for the proposed algorithms will be focused.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Statistical-based methods to improve precision of DBSCAN clustering algorithm for obstacle detection application in autonomous vehicles

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月16日

查看全文
  • Harn Tung Ng,
  • Haidi Ibrahim,
  • Parvathy Rajendran

摘要

Recently, mmWave Radar sensor has been used widely as perception sensor in autonomous vehicles due to its compact size and light weight. Typically, the mmWave Radar point cloud data generated are clustered using algorithms like Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, to associate the point cloud data to each of the obstacles present in the environment. However, the parameters for the DBSCAN algorithm are often fixed, and it does not take into account the varying environment encountered by the autonomous vehicles. Furthermore, the DBSCAN algorithm does not consider the radial velocity of each data point throughout the clustering process. All these problems can indeed lead to less precise clustering performance. Hence, this paper presented two statistical-based approaches to improve the precision of DBSCAN algorithm. The proposed K-Nearest Neighbors - Probability (KNN-Probability) method determines the eps parameter of DBSCAN algorithm dynamically for each point cloud data frame, while the proposed cluster refinement algorithm refines the DBSCAN clusters based on radial velocity data. By utilizing the publicly available RadarScenes dataset for performance evaluation, the proposed obstacle detection algorithm has achieved a more precise clustering result with a high mean average precision (mAP) of 0.8942. In terms of future research direction, the development of artificial intelligence (AI) based methods to optimize the selection of parameters for the proposed algorithms will be focused.