Application of Continuous Learning Algorithm in Vehicle Target Recognition

doi:10.16638/j.cnki.1671-7988.2023.015.013

Abstract

Abstract: Advances in deep learning and artificial intelligence are the main drivers of autonomous vehicle technology. However, most of the deep learning models are trained on the same staticdistributed data set and their behaviors can not be adaptde or expended over time. To solve this problem, the continuous learning model is applied in the field of vehicle object recognition. Firstly, an environment is built to make the algorithm run smoothly, and the original image data set of target recognition is selected. On the basis of analyzing the existing evaluation indexes, the evaluation indexes suitable for this experiment are selected, and convolutional neural network (CNN), nearest class mean (NCM), incremental classifier and representation learning (iCaRL) three continuous learning algorithms are used for learning, training and comparing the original image data set. Based on the experiment result, the iCaRL algorithm makes the accuracy and efficiency of machine continuous learning training better than the other two methods. Aiming at the problem that the image data set of intelligent driving target recognition is not perfect, a new image data set is constructed, including vehicles, pedestrians, traffic signs and signal lights. ICaRL algorithm is applied to the new image data set for research, and the new intelligent draving image dataset is trained and tested. The results show that iCaRL algorithm can learn the new image data in the environment. The test results show that the method can be used for target recognition in intelligent driving field.

Key words: Continual learning; ICaRL algorithm; Vehicle target recognition; Image dataset

摘要： 自动驾驶汽车技术的日新月异，主要得益于深度学习和人工智能的进步。然而深度学习模型大多是在静态同分布数据集上进行训练，无法随着时间而适应或扩展其行为。针对这一问题，论文将持续学习模型运用于车辆目标识别领域进行研究。首先搭建可以使得算法流畅运行的环境，选定目标识别的原始图像数据集；在分析现有评估指标的基础上，选取适合于本次实验的评估指标，并采用卷积神经网络（CNN）、最接近类均值（NCM）、增量分类器与特征表示（iCaRL）三种持续学习算法对原始图像数据集进行学习训练与对比验证，通过实验验证了应用 iCaRL 算法使机器进行持续学习训练时，其精度和效率均优于其他两种方法。针对智能驾驶目标识别图像数据集不完善这一问题，构建了一个新的图像数据集，包含车辆、行人、交通标志及信号灯，将 iCaRL 算法应用于新建图像数据集进行研究，并在新建智能驾驶图像数据集上进行了训练与测试。结果表明，采用 iCaRL 算法能够较好地学习新建图像数据集，不会因为环境的改变而使得其性能发生大幅变化，测试结果良好，证明该方法可以在智能驾驶领域进行目标识别。

关键词: 持续学习；iCaRL 算法；车辆目标识别；图像数据集

SUN Jiahui. Application of Continuous Learning Algorithm in Vehicle Target Recognition[J]. Automobile Applied Technology, 2023, 48(15): 73-81.

孙家辉. 持续学习算法在车辆目标识别上的应用[J]. 汽车实用技术, 2023, 48(15): 73-81.

References

[ 1 ] 韩亚楠,刘建伟,罗雄麟.连续学习研究进展[J].计算机研究与发展,2022,59(6):1213-1239. [ 2 ] 丁威.智能辅助驾驶场景典型目标识别技术研究[D]. 成都:电子科技大学,2020. [ 3 ] 刘建军,胡卫东,郁文贤.一种 RBF 网络结构调整的稳健增量学习方法[J].计算机仿真,2009,26(7):192- 194,227. [ 4 ] 陈焕文.网络分割与自适应突触可塑性融合的连续学习[J/OL].华中科技大学学报(自然科学版):1-7 [2023-08-09].http://doi.org/10.13245/j.hust.230052. [ 5 ] 党思航.目标识别中的增量学习方法研究[D].成都: 电子科技大学,2021. [ 6 ] REBUFFI S A,KOLESNIKOV A,SPERL G,et al. Incremental Classifier and Representation Learning [C]//Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE,2017:5533-5542. [ 7 ] 郑远攀,李广阳,李晔.深度学习在图像识别中的应用研究综述[J].计算机工程与应用,2019,55(12):20-36. [ 8 ] 刘禹希.基于机器视觉的智能驾驶车辆的目标识别研究[D].长春:吉林大学,2017. [ 9 ] 罗雨珊.基于 cnn 的车辆目标检测关键技术研究与设计[D].成都:电子科技大学,2019. [10] CAI Z,FAN Q,FERIS R S,et al.A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection[C]//14th European Conference. Heidelberg: Springer,2016:354-370. [11] GUERRIERO S,CAPUTO B,MENSINK T.DeepNCM: Deep Nearest Class Mean Classifiers[C]//International Conference on Learning Representations.Vancouver: ICLR,2018. [12] KANG B,XIE S,ROHRBACH M,et al.Decoupling Representation and Classifier for Long-tailed Recognition[J].arXiv Preprint arXiv:1910.09217,2019. [13] PARK J,KANG M,HAN B.Class-incremental Learning for Action Recognition in Videos[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.Piscataway:IEEE,2021:13698-13707. [14] VAN DE VEN G M, TOLIAS A S.Three Scenarios for Continual Learning[J].arXiv Preprint arXiv:1904.07734, 2019. [15] 张璨,张明英.人工智能深度学习算法可靠性评估方法研究[J].信息技术与标准化,2018(8):38-42. [16] 吴楚,王士同.任务相似度引导的渐进深度神经网络及其学习[J].计算机科学与探索,2023,17(5):1126- 1138. [17] 王强,曹磊,史润佳,等.基于 DCResNet 的 SAR 图像车辆目标识别[J].雷达科学与技术,2021,19(4):387-392, 402. [18] 王庆伟.基于数据预处理与卷积神经网络的调制信号识别研究[D].南京:南京信息工程大学,2022. [19] CHAUDHRY A, RANZATO M A, ROHRBACH M, et al. Efficient Lifelong Learning with A-gem[J].arXiv Preprint arXiv:1812.00420, 2018. [20] 郭阅.基于深度学习的智能车前方目标识别与预警研究[D].长春:吉林大学,2020. [21] 胡艳,曹钻.基于机器视觉和深度学习的自动驾驶方法研究[J].环境技术,2021,39(3):100-105,110.