摘要
介绍了目标检测数据集的发展过程、基本评价指标的设定,并基于此综述了不同类别的目标检测算法,分别对两阶段和单阶段检测算法及相应优化算法进行解析,围绕检测速度和检测精度的迭代过程,阐述了目标检测算法的困难与挑战。最后,就算法本身的提升和算法应用需求下的优化设计提出总结和展望,指出目标检测的训练监督问题、算法对小目标的检测困难问题,同时指出实时检测任务中检测速度与检测精度的协调性问题和多模态融合应用问题,以及算法运行可解释性对算法再提升的重要意义。
This paper introduced the development of object detection datasets and the establishment of basic evaluation metrics,and based on this,it reviewed different categories of object detection algorithms.Single-stage and two-stage detection algorithms,as well as corresponding optimization algorithms,were analyzed separately.Highlighting the iterative process of detection speed and accuracy,the paper elaborated the challenges and difficulties in object detection algorithms.A summary and outlook for the improvement of the method itself and the optimization design under the application requirements of the algorithm were proposed in the paper,which indicated training supervision of object detection,the difficulty of detecting small targets by the algorithm.At the same time,the paper also indicated the coordination between detection speed and accuracy in real-time detection tasks and multimodal fusion application,as well as the important significance of the interpretability of algorithm operation for further improving the algorithm.
作者
曾文炳
李军
Zeng Wenbing;Li Jun(Chongqing Jiaotong University,Chongqing 400074)
出处
《汽车工程师》
2024年第1期1-11,共11页
Automotive Engineer
基金
重庆市研究生联合培养基地项目(JDLHPYJD2018003)。
关键词
目标检测算法
深度学习
计算机视觉
卷积神经网络
Target detection algorithm
Deep learning
Computer vision
Convolution neural network
