基于EdgeBoxes与旋转不变特征的车辆检测

Car detection based on EdgeBoxes and rotation-invariant features

为能在车辆姿态变化较大时高效准确地检测出车辆,提出一种基于Edge Boxes与旋转不变特征的车辆检测算法。通过Edge Boxes算法,确定车辆候选区域(proposals),然后计算图像局部敏感直方图,提取光照不变特征;之后计算基于快速傅里叶变换的傅里叶梯度方向直方图(Fourier HOG),提取旋转不变特征;得到特征描述符(descriptor)后,使用线性支持向量机(SVM)进行训练分类。实验结果表明,该算法在提高了检测准确率的同时也提升了检测速度。

To detect car efficiently and accurately when car posture gets seriously changed,a car detection algorithm based on EdgeBoxes and Rotation-Invariant Features is proposed. By EdgeBoxes algorithm,car proposals can be obtained via car edges. By calculating the Locality Sensitive Histograms,Illumination Invariant Features are extracted. Then the Fourier Histogram of Oriented Gradients based on Fast Fourier Transform is calculated,Rotation-Invariant Features are extracted. As the feature descriptor is obtained,linear SVM is used to train and classify.Experimental results show that,the algorithm improves the precision of detection and becomes more efficient.