车载传感器数据与高精度地图的精确匹配是提升智能车辆感知与定位的关键.提出基于语义似然模型(SLM)的高精度地图匹配新算法,实现智能车辆同时定位与目标检测任务.首先通过U-Net提取路面语义目标,利用核密度估计建立SLM.基于粒子滤波框架,利用位姿变换将高精度地图上目标抽样点映射至SLM中,计算该抽样点与传感器数据的匹配程度对每个粒子的权重更新,实现智能车辆的高精度定位.最后利用定位结果完成地图上的要素目标到图像的映射,实现目标的精准检测.利用在校园道路与城市道路环境下采集的数据对所提算法进行验证,实验结果表明,算法的平均定位误差约为14 cm,路面路标检测结果平均交并比(MIoU)均大于80.较之深度神经网络等当前最佳(SOTA)的检测方法,所提算法引入高精度地图的先验信息可显著提升智能车辆定位与目标检测性能.

Accurate matching between in-vehicle sensor data and high-definition (HD) maps is crucial to improve the performance of perception and localization of intelligent vehicles. A novel algorithm of HD map matching based on the developed semantic likelihood model (SLM) is proposed to achieve intelligent vehicle localization and object detection simultaneously. First, semantic pavement objects are extracted from front-view images by using U-Net, and SLM is constructed with kernel density estimation (KDE). Under a particle filter framework, the likelihood between the sensor data and HD map is calculated by projecting each sample point from HD map with pose transformation onto SLM to update the weight of each particle. Simultaneously, accurate detection of pavement markings is accomplished by projecting all elements onto the HD map with the computed localization results. In the experiment, data collected on campus and on urban roads are used to validate the proposed algorithm. The experimental results show that the localization errors in both scenarios are about 14 cm, and the mean intersection over union (MIoU) of road marking detection is above 80. The results demonstrate that the proposed algorithm can significantly improve both localization and detection performance by effectively utilizing the prior information of HD maps, compared with the state of the art (SOTA) methods, such as deep learning-based detection methods.