基于改进长短时记忆神经网络-自适应增强算法的多天气车辆分类方法

Vehicle Classification Method in Multi-climates Based on Modified LSTM-AdaBoost Algorithm

针对目前国内外车辆分类效果不理想和受天气影响较大的问题,本文中提出一种基于改进长短时记忆神经网络自适应增强算法(LSTM-AdaBoost)的多天气车辆分类方法,并提出一种“多层网格法”以准确地确定LSTM的超参数。首先建立地磁车辆检测系统平台和车辆分类方法,然后分析基于改进LSTM-AdaBoost的车辆分类结果,并对不同车辆分类方法和不同天气下的分类准确率进行了对比。结果表明,与最邻近结点算法和反向传播神经网络算法相比,本文所提出的方法具有较高的准确率,最高分类准确率为92.2%。暴雨、雾霾和晴天3种天气中,暴雨时的分类准确率最低,但差别不大,最大相差3.9个百分点。

In view of the poor results of existing domestic and oversea vehicle classification schemes and relatively significant effects of climate on them,a multi-climate vehicle classification method based on modified LSTM-AdaBoost(long short-term memory neural network-Adaptive boosting)algorithm is proposed,and a“multi-layer grid method”is also put forward to accurately determine the hyperparameters of LSTM.Firstly,the geomagnetic vehicle detection system and vehicle classification method are established.Then the results of vehicle classification based on modified LSTM-AdaBoost are analyzed,and the classification accuracies of different vehicle classification methods and different climates are compared.The results show that compared with K-nearest neighbor and BP neural network algorithms for classification,the proposed method has higher accuracy with a highest classification accuracy of 92.2%.Among three climates of torrential rain,haze and fine day,the classification accuracy in torrential rain is lowest,but the difference is rather small,3.9 percentage points at most.