融合SIFT和级联分类器的特种车辆自动检测识别

Automatic Detection and Recognition of Special Vehicles Incorporating SIFT and Cascade Classifier

针对特定场景中特种车辆因多环境影响因素下数据不均衡、检测精度和识别准确率低的问题,提出一种融合尺度不变特征变换(Scale-Invariant Feature Transform,SIFT)和级联分类器的特种车辆自动检测及识别预测方法。首先,图像预处理后运用SIFT特征提取图像主体区域特征点及特征描述子;其次,结合SIFT特征点的密度调整优化算法实现目标车辆检测;最后,运用KMeans聚类算法获得目标检测框中SIFT特征描述子的中心聚类点,生成表征目标主体图像的128维特征描述子,并最终输入RF-RBF(Random Forest-Radial Basis Function)级联分类器进行学习并识别预测。该文均采用K折交叉验证方法保证模型的稳定性和可靠性。实验结果表明,在特定场景下特种车辆目标检测获得了75.47%平均交并比,级联分类器在特种车辆识别的综合准确率、精确率、召回率、F1-Score值以及FPS值分别为87.35%、88.17%、97.27%、92.38%以及21。进一步验证融合SIFT和级联分类器模型具有较好的自动化检测准确性和识别分类能力。

In order to overcome the unbalanced data and low detection and recognition accuracy of special vehicles in specific scenes caused by a number of environmental influencing factors,a Scale-Invariant Feature Transform(SIFT)and cascade classifier method for special vehicle identification and recognition prediction is proposed.Firstly,the SIFT features are used to extract the feature points and feature descriptors of the image subject area after image pre-processing.Secondly,the density adjustment optimization algorithm of SIFT feature points is combined to realize the target vehicle detection.Finally,the KMeans clustering algorithm is used to obtain the central clustering points of the SIFT feature descriptors in the target detection frame to generate 128-dimensional feature descriptors characterizing the target subject image,and finally input RF-RBF(Random Forest-Radial Basis Function)cascade classifier for learning,identifying and prediction.All K-fold cross-validation methods are used to ensure the stability and reliability of the model.The experimental results show that 75.47%average cross-comparison ratio is obtained for special vehicle target detection in specific scenarios,and the combined accuracy,precision,recall,F1-Score and FPS values of the cascade classifier in special vehicle recognition are 87.35%,88.17%,97.27%,92.38%and 21,respectively.Such model has better automatic detection accuracy and recognition classification ability.