汽车工程 ›› 2021, Vol. 43 ›› Issue (1): 50-58.doi: 10.19562/j.chinasae.qcgc.2021.01.007

• • 上一篇    下一篇

基于类人视觉的多任务交通目标实时检测模型

刘军(),陈岚磊,李汉冰   

  1. 江苏大学汽车与交通工程学院,镇江 212013
  • 收稿日期:2020-06-01 修回日期:2020-08-08 出版日期:2021-01-25 发布日期:2021-02-03
  • 通讯作者: 刘军 E-mail:Liujun@ujs.edu.cn
  • 基金资助:
    国家自然科学基金(51275212)

A Real⁃time Detection Model for Multi⁃task Traffic Objects Based on Humanoid Vision

Jun Liu(),Lanlei Li Hanbing Chen   

  1. School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang 212013
  • Received:2020-06-01 Revised:2020-08-08 Online:2021-01-25 Published:2021-02-03
  • Contact: Jun Liu E-mail:Liujun@ujs.edu.cn

摘要:

针对复杂交通场景下单个模型无法同时对可行驶区域和交通目标进行统一检测和检测实时性较差的问题,本文中提出了一种基于类人视觉的多任务交通目标单阶段检测模型,实现了可行驶区域和车辆、行人的实时统一检测。首先建立类人视觉注意力机制进行监督,采用轻量化模型MobileNetV3作为骨架,接着利用特征金字塔的思想对可行驶区域进行检测,最后采用动态注意力算法和anchor?free思想对交通目标进行检测。实验结果表明,引入驾驶员人眼注意力机制明显提升了模型的精度和鲁棒性,模型在实车上平均运算速度达到了12帧/s。

关键词: 类人视觉, 注意力机制, anchor?free, 交通场景检测

Abstract:

In view of that in complex traffic scenes the single model can't carry out unified detection of drivable region and traffic objects concurrently and the poor real?time performance of detection, a single?stage detection model for multi?task traffic object based on humanoid vision is proposed in this paper, achieving the unified detection of drivable region, vehicle and pedestrian concurrently in real time.Firstly, a humanoid vision attention mechanism is established for supervision and lightweight model MobileNetV3 is adopted as a framework. Then the idea of feature pyramid network (FPN) is utilized to detect drivable region. Finally, dynamic attention algorithm and the idea of anchor?free are used to detect traffic object. The experimental results show that the attention mechanism of driver's eyes introduced obviously enhance the accuracy and robustness of the model and the average calculation speed of the model in real vehicle reaches 12 f/s.

Key words: humanoid vision, attention mechanism, anchor?free, traffic scene detection