Scholarlink

<p>Object detection using point cloud data in three-dimensional space is garnering increasing attention in the field of intelligent vehicles. Given LiDAR data possesses characteristics of sparseness and uneven distribution, missed detections and false positives still occur intermittently. Therefore, this paper explores the structure of the detection network to ameliorate this issue. Initially, this study employs PointPillars, a single-stage network, as the foundational baseline and subsequently develops RAK-Pillars, a sophisticated three-dimensional point cloud object detection algorithm predicated on residual networks and adaptive kernel convolution. The network’s feature extraction proficiency is augmented through the construction of identity mapping, while AKConv supplants the conventional convolution layer, enabling the network to employ convolution kernels with diverse parameters to execute convolution operations of various shapes and sizes. Empirical results reveal that the detection performance of the RAK-Pillars algorithm markedly exceeds that of the baseline network, particularly in mitigating false positives and missed detections of pedestrians and cyclists. Moreover, the deep learning detector algorithm is further examined on the RAM-Pillars foundation. In the original detection head of the PointPillars network, a solitary detector concurrently learns the features of vehicles, bicycles, and pedestrians, without addressing the considerable differences in their size information. This paper introduces a multi-detector framework, utilizing distinct detectors for each category to enhance detection accuracy and incorporating an efficient channel attention mechanism to further refine the precision of the multi-category multi-detection heads. Experimental results corroborate that the detection efficacy of the RAM-Pillars algorithm is further enhanced, with improved detection capabilities for automotive objects, while maintaining the accuracy for pedestrians and cyclists. Real-world vehicle data collected on campus were employed for experimental validation, demonstrating that this approach sustains high detection accuracy while fulfilling real-time processing requirements.</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

Research on Laser Point Cloud 3D Object Detection Based on Deep Learning

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • yibing Zhao,
  • Zaifa Ding,
  • Yuqiao Chen,
  • Lie Guo

摘要

Object detection using point cloud data in three-dimensional space is garnering increasing attention in the field of intelligent vehicles. Given LiDAR data possesses characteristics of sparseness and uneven distribution, missed detections and false positives still occur intermittently. Therefore, this paper explores the structure of the detection network to ameliorate this issue. Initially, this study employs PointPillars, a single-stage network, as the foundational baseline and subsequently develops RAK-Pillars, a sophisticated three-dimensional point cloud object detection algorithm predicated on residual networks and adaptive kernel convolution. The network’s feature extraction proficiency is augmented through the construction of identity mapping, while AKConv supplants the conventional convolution layer, enabling the network to employ convolution kernels with diverse parameters to execute convolution operations of various shapes and sizes. Empirical results reveal that the detection performance of the RAK-Pillars algorithm markedly exceeds that of the baseline network, particularly in mitigating false positives and missed detections of pedestrians and cyclists. Moreover, the deep learning detector algorithm is further examined on the RAM-Pillars foundation. In the original detection head of the PointPillars network, a solitary detector concurrently learns the features of vehicles, bicycles, and pedestrians, without addressing the considerable differences in their size information. This paper introduces a multi-detector framework, utilizing distinct detectors for each category to enhance detection accuracy and incorporating an efficient channel attention mechanism to further refine the precision of the multi-category multi-detection heads. Experimental results corroborate that the detection efficacy of the RAM-Pillars algorithm is further enhanced, with improved detection capabilities for automotive objects, while maintaining the accuracy for pedestrians and cyclists. Real-world vehicle data collected on campus were employed for experimental validation, demonstrating that this approach sustains high detection accuracy while fulfilling real-time processing requirements.