Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (6): 183-191.DOI: 10.3778/j.issn.1002-8331.2109-0190

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Object Detection in Autonomous Driving Scene Based on Improved Efficientdet

LI Yanchen, ZHANG Xiaojun, ZHANG Minglu, SHEN Liangyi   

  1. College of Mechanical Engineering, Hebei University of Technology, Tianjin 300401, China
  • Online:2022-03-15 Published:2022-03-15

基于改进Efficientdet的自动驾驶场景目标检测

李彦辰,张小俊,张明路,沈亮屹   

  1. 河北工业大学 机械工程学院,天津 300401

Abstract: Aiming at the problems of limited computing resources on the vehicle-mounted platform and low detection accuracy of small targets in autonomous driving scenarios, a single-stage object detection framework Efficientdet-Gs based on Efficientdet is proposed. The backbone network Efficientnet is improved by reconstructing the inverted residual bottleneck MBConv, which reduces the amount of network parameters and calculations without sacrificing accuracy. The multi-scale attention mechanism module is designed to be applied to the feature fusion network, which further improves the detection accuracy of small targets. The Balanced L1 Loss is introduced to replace the original regression loss function Smooth L1 Loss, which solves the problem of balance in the loss function. Experimental results show that, compared with Efficientdet, the calculation of Efficientdet-Gs is reduced by an average of 25%, the average detection accuracy on the BDD100K test set is increased by 4.8%, and the average inference speed is increased by 5.7%. This framework can achieve good detection results when the hardware requirements of vehicle-mounted equipment are low.

Key words: autonomous driving, object detection, deep learning, Efficientdet, Ghostnet, attention mechanism

摘要: 针对自动驾驶场景中车载平台计算资源有限及小目标检测精度较低等问题,提出一种基于Efficientdet的单阶段目标检测框架Efficientdet-Gs。通过重构倒转残差瓶颈MBConv来改进主干网络Efficientnet,在不牺牲精度的同时降低了网络的参数量和计算量;设计多尺度注意力机制模块应用于特征融合网络,进一步提高了对小目标的检测精度;引入Balanced L1 Loss替换原回归损失函数Smooth L1 Loss,解决了损失函数中的平衡性问题。实验结果表明,Efficientdet-Gs的平均计算量相较于Efficientdet下降了25%,在BDD100K测试集上平均检测精度提高了4.8%,平均推理速度提高了5.7%。该框架在车载硬件设备要求较低的情况下能够实现良好的检测效果。

关键词: 自动驾驶, 目标检测, 深度学习, Efficientdet, Ghostnet, 注意力机制