一种基于可变形部件模型的快速对象检测算法

摘要

To solve the speed bottleneck of deformable part models in the detection process, this paper proposes a cascade deformable part model with rapid computation of feature pyramids for the detection process of the model. Because the speed of the detection is mainly determined by the two processes of the feature computation and the object location, a two-stage speedup algorithm is proposed. Firstly, sparsely-sampled feature pyramids on the scale are utilized to approximate finely-sampled multi-scale image features to speed up the process of feature computation. Then combined with the cascade algorithm in the location process, a sequence model is utilized to evaluate individual parts sequentially so as to rapidly prune most object hypotheses of small possibilities in order to speed up the process of object location. The experimental results on PASCAL VOC 2007 dataset and INRIA dataset show that the algorithm in the paper apparently speeds up the speed of detection with minor loss in detection precision.%为了解决可变形部件模型检测过程中的速度瓶颈问题,该文针对模型的检测流程,提出一种结合快速特征金字塔计算的级联可变形部件模型.由于模型的检测速度主要取决于特征计算以及对象定位这两个过程,提出一种两阶段的加速算法:首先采用尺度上稀疏采样的特征金字塔来近似表示精细采样的多尺度图像特征,以加快特征计算过程;然后在定位过程中结合级联算法,以一个序列模型顺序地评估各个部件,从而快速剪除大部分可能性较小的对象假设,以加快对象定位过程.在PASCAL VOC 2007和INRIA数据集上的实验结果表明,该算法可以明显加快检测速度,而检测精度仅略有下降.