汽车工程 ›› 2023, Vol. 45 ›› Issue (4): 541-550.doi: 10.19562/j.chinasae.qcgc.2023.04.002

所属专题: 智能网联汽车技术专题-感知&HMI&测评2023年

• • 上一篇    下一篇

基于检测的高效自动驾驶实例分割方法

陈妍妍1,王海1(),蔡英凤2,陈龙2,李祎承2   

  1. 1.江苏大学汽车与交通工程学院,镇江   212013
    2.江苏大学汽车工程研究院,镇江  212013
  • 收稿日期:2022-11-11 修回日期:2022-11-30 出版日期:2023-04-25 发布日期:2023-04-19
  • 通讯作者: 王海 E-mail:wanghai1019@163.com
  • 基金资助:
    国家自然科学基金(52225212);江苏省重点研发项目(BE2020083-2)

Efficient Automatic Driving Instance Segmentation Method Based on Detection

Yanyan Chen1,Hai Wang1(),Yingfeng Cai2,Long Chen2,Yicheng Li2   

  1. 1.School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang  212013
    2.Institute of Automotive Engineering,Jiangsu University,Zhenjiang  212013
  • Received:2022-11-11 Revised:2022-11-30 Online:2023-04-25 Published:2023-04-19
  • Contact: Hai Wang E-mail:wanghai1019@163.com

摘要:

基于深度学习的实例分割算法在大规模通用场景中取得了良好的分割性能,然而面向复杂交通场景的多目标实例分割仍然极具挑战性,尤其在算法的高精度和较快推理速度的权衡方面,而这对于智能汽车的行驶安全性至关重要。鉴于此,本文以实时性算法Orienmask为基础,提出了一种基于单阶段检测算法的多头实例分割框架。具体来说,所提框架由骨干网络、特征融合模块和多头掩码构建模块组成。首先,本文通过在骨干网络中加入残差结构获取更加完整的高维表征信息。其次,为了产生更具判别性的特征表达,本文通过引入自校正卷积重构特征金字塔,并使用全局注意力机制改善信息传播以进一步优化所提框架的特征融合模块。最后,提出的多头掩码构建机制通过细化场景目标尺寸分布显著提高不同目标的分割性能。本文算法在开源数据集BDD100k上进行大量测试与验证,分别在边界框和掩码上获得了23.3% 和19.4%的均交并比(mAP@0.5:0.95),与基线方法相比,平均指标提高了5.2%和2.2%。同时在基于自主搭建的实车平台上进行的道路实验也证明本算法能够较好地适应真实驾驶环境,且满足实时性分割需求。

关键词: 自动驾驶, 深度学习, 目标检测, 实例分割

Abstract:

The instance segmentation algorithm based on deep learning has achieved excellent performance in large-scale general scenarios. However, the segmentation of multi-objective instances for complex traffic scenes is still challenging, especially in the balance between high accuracy and fast inference speed, which is crucial to driving safety of intelligent vehicles. In view of this, based on the real-time algorithm Orienmask, a multi-head segmentation framework is proposed based on the one-stage detection method. Specifically, the proposed framework comprises of a backbone, a feature fusion module and a multi-head mask construction module. Firstly, complete high-dimensional feature maps are obtained by adding residual structures to the backbone.Secondly, in order to generate discriminative feature representations, the feature pyramid module is reconstructed by introducing in self-calibrate convolutions and the information propagation path is improved by global attention mechanism, so as to further optimize the feature fusion module of the proposed framework. Finally, a multi-head mask construction mechanism is proposed to significantly improve the segmentation performance of different targets by refining the size distribution of instances in the traffic scenes. The proposed algorithm has been tested and validated on the open-source dataset BDD100k, and has achieved an average intersection ratio of 23.3% and 19.4% (mAP@0.5:0.95) on bounding boxes and segmentation masks, respectively. Compared with the baseline, the average index are increased by 5.2 % and 2.2 %. At the same time, the road experiment on the self-built real-vehicle platform also proves that the proposed algorithm can adapt to actual driving environments and meet the demands of real-time segmentations.

Key words: autonomous vehicles, deep Learning, object detection, instance segmentation