基于改进的DSSD算法的行人检测  

Pedestrian Detection Based on Improved DSSD Algorithm

作　　者：邝先验 杨江波 张建华 Kuang Xianyan;Yang Jiangbo;Zhang Jianhua(School of Electrical Engineering and Automation,Jiangxi University of Science and Tecnology,Ganzhou,Jiangxi 341000)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341000

出　　处：《中国仪器仪表》2021年第5期21-27,共7页China Instrumentation

基　　金：国家自然科学基金(51268017,61463020);江西省教育厅科技项目(GJJ160609)。

摘　　要：交通场景下的行人检测在高级辅助驾驶系统和自动驾驶汽车领域中占有重要地位。为了解决道路行人因采集视角和低像素模糊而导致小尺度行人低检测精度的问题,提出了一种基于DSSD的行人检测网络框架,结合改进的ResNeXt特征提取模型作为DSSD检测框架的前置网络以保证小尺度行人特征的精确提取和高效传递。为了充分获取局部细节信息和全局语义信息,通过对深层网络进行反卷积操作提取不同尺度的特征,并将其通过采用相同的FPN融合策略与浅层网络进行多尺度的特征融合。最后在INRIA数据集上进行训练和测试,实验结果表明,与其他算法相比该方法具有更高的准确率和召回率。Pedestrian detection in traffic scenes plays an important role in the field of advanced auxiliary driving system and autopilot.In order to solve the problem of low detection accuracy of smallscale pedestrian due to the collection perspective and low pixel blur,a pedestrian detection network framework based on dssd is proposed.Combined with the improved resnext feature extraction model,it is used as the front network of dssd detection framework to ensure the accurate extraction and efficient transmission of small-scale pedestrian features.In order to fully obtain local details and global semantic information,different scale features are extracted by deconvolution of deep network,and multi-scale features are fused with shallow network by using the same FPN fusion strategy.Finally,the algorithm is trained and tested on INRIA data set,and the experimental results show that the algorithm has higher accuracy and recall rate than other algorithms.

关 键 词：行人检测 DSSD检测模型 特征融合 ResNeXt特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U463.6[自动化与计算机技术—计算机科学与技术]