Abstract: Aiming at the problem that running slowly when using deformable part model to detect vehicles due to its high complexity, an improved deformable part model is proposed. On the one hand, weighted PCA is used to reduce dimension of HOG feature, the basis of deformable part model, so that the model’s parameters can be decreased. On the other hand, after the combination of HOG feature-levels, Fast Fourier Transform（FFT） is applied to convert convolution between filters and HOG feature-levels into multiplication in frequency domain, which reduces the computation complexity. The experimental results show that the improved deformable part model gains similar precision and recall-rate compared with original model, while its speed increases significantly, whose average consuming time accounts for 29.6% and 26.3% of original model respectively on UIUC dataset and BIT dataset.

Key words: deformable part model, vehicle detection, Weighted Principal Component Analysis（WPCA）, feature-levels combination

摘要： 针对可变形部件模型的复杂性使其在检测车辆时速度慢的问题，对可变形部件模型进行了改进。一方面使用加权PCA对可变形部件模型的基础HOG特征进行降维来减少模型参数；另一方面将HOG特征层组合后，使用快速傅里叶变换（FFT）把滤波器与HOG特征层的卷积转换为频域乘积，来降低计算复杂度。仿真实验结果表明，改进的可变形部件模型在进行车辆检测时检测精度和召回率都与原始模型相当，但检测速度大幅提升，在UIUC和BIT两个数据集上的平均耗时分别仅占原始模型平均耗时的29.6%和26.3%。

关键词: 可变形部件模型, 车辆检测, 加权主成分分析, 特征层组合