特征级联在车辆检测中的应用与研究

Application and Research of Feature-Cascade in Vehicle Detection[JZ)][HS)]

摘要:

对现有的采用机器学习算法检测车辆进行研究,分析其存在的不足；表现在特征或者算法单一,对光照等条件变化鲁棒性不够；针对这些问题,提出一种融合LBP特征与HOG特征,并结合Adaboost与SVM的车辆检测算法；借鉴级联的思想,首先采用AdaBoost对训练样本提取LBP特征进行训练,得到的分类器用于初步筛选并将其分类结果作为下一层分类器的输入；然后采用SVM算法对训练样本提取HOG特征进行训练,得到的分类器用于二次筛选上一层分类器的分类结果；实验结果证明Adaboost-SVM相结合的办法检测结果精度高,准确率和召回率均达到95%以上,FPPW与FPPI的值均在5%左右；同时由于算法采用的特征对光照条件具有较强的鲁棒性,因此光照条件的变化对算法的识别结果影响较低；实时性方面,每帧图像的处理时间为75 ms,满足实时性要求。

Abstract:

Do researches on the existing machine learning algorithms of vehicles detecting and analyze the disadvantages, which lie in using single feature or single ML algorithms, or not robust enough of illumination and other condition changes. For these problems, a vehicle detecting algorithm is presented based on fusion of LBP and HOG features and combining the Adaboost and SVM. Referencing cascade thought, the algorithm first extracts LBP features of training samples by AdaBoost and the classifier is used for initial classification. Classification results act as the inputs of next classifier. Then HOG features of training samples are extracted by SVM and the resulting classifier is applied to classify the above results. Experiments show that AdaBoost-SVM method accuracy rate and recall rate both are above 95% and the value FPPW of FPPI are about 5%. At the same time the proposed method is robust to lighting conditions because of the selected features. In the aspects of real-time, each frame processing time is 75ms which meets the real-time requirements.