摘要
针对在行人检测领域,深度学习的方法需在GPU下运算,限制了实用性;机器学习的硬件要求低,但其实时性差的问题,提出了一种基于DPM的改进算法——F-DPM算法。F-DPM对DPM进行了3项改进:(1)通过对soft binning直方图降维快速构建特征金字塔;(2)通过FFT加快了特征与模板的卷积运算速度;(3)通过分层检测保证检测目标位置的准确性。改进后的F-DPM算法的检测准确率为85.7%,召回率为82.9%,平均检测速度为52 ms,基本满足实时性和准确性要求。对比ACF、YOLO算法,其复杂场景的检测效果更优,这为行人检测的实用化起到了一定推进作用。
Requiring the assistance of GPU restricts the practical application of deep learning method in the area of pedestrian detection. The real-time performance is poor,if low-level configuration of hardware is used. To cope with it,an improved algorithm based on the DPM,which is named F-DPM algorithm,is proposed in this paper. The F-DPM makes three improvements to DPM:(1) the feature pyramid is built quickly by reducing the dimension of soft binning histogram;(2) the convolution operation of feature and template is accelerated by FFT;(3) the accuracy of the target location is guaranteed by layer detection. Final results show that the precision of F-DPM algorithm reaches 85. 7%,the recall of the algorithm is 82. 9%,and the average speed of detection is 52 ms. Basically,it meets the real-time and accuracy requirements. Comparing with the ACF and YOLO algorithms,the test results of F-DPM is better in complex scene. Therefore,this F-DPM algorithm contributes to the practical application in detecting pedestrians.
