摘要： 为提高行人检测的准确性与稳定性,提出一种新的检测方法。以SSD方法为基础进行优化与改进网络结构,将串联式的基础网络部分修改为密集连接式结构,在目标预测阶段选择融合特征作为预测依据,根据目标尺寸的统计分布规律调整不同特征层的缩放因子。在Caltech数据集上的测试结果表明,相比于原始SSD、VJ-1、HOG等方法,该方法具有更高的准确性和更好的鲁棒性,尤其对于待检行人目标尺寸较小以及严重遮挡等行人检测问题,改进SSD方法检测结果更好。该方法在TitanX测试条件下具有20 frame/s的检测速度,满足实时性要求。

关键词: 行人检测, 卷积神经网络, 融合特征, 密集连接, 多尺度检测

Abstract: To improve accuracy and efficiency of pedestrian detection,this paper proposes a method for pedestrian detection based on modified Single Shot Multibox Detector (SSD).The network model of this method is modified and optimized according to the SSD method.The tandem base network is modified to a densely connected structure to improve the expressiveness of the model.In order to improve the robustness,the fusion feature is selected as the prediction basis.The scale ratios of different feature layers are also adjusted according to statistic distribution rule of pedestrian size to improve prediction accuracy.The model is trained and tested via Caltech Pedestrian dataset.Experimental results show that,compared with the original SSD and other state-of-the-art methods,the propsed method has higher accuracy and better robustness for pedestrian detection problems.Especially for pedestrians with small target size and heavy occlusion,the propsed method has a more significant improvement in detection results.The model is also tested via TitanX,which achieves a detection speed as high as 20 frame/s.It means that it can used in the real-time scenes.

Key words: pedestrian detection, Convolutional Neural Network(CNN), fusion feature, dense connection, multi-scale detection

中图分类号: 

• TP391.4