基于多尺度特征融合的小尺度行人检测

Small-scale Pedestrian Detection Based on Multi-scale Feature Fusion

随着无人驾驶技术的蓬勃发展,针对行人的检测成为一大难点,同时也是热点研究问题。而针对传统行人检测框架(One-stage和Two-stage等)对小尺度行人检测效果不佳的问题,本文在FPN网络基础上尝试了新的策略,致力于提高视频序列不同尺度行人的识别精度。算法先通过ResNet50提取特征,并采用FPN进行多尺度特征融合,同时利用RPN产生推荐区域,最后Fast RCNN对RPN产生的推荐区域实现分类与回归,经过非极大值抑制后处理等到最终结果。实验结果表明,本文基于FPN构建的行人检测算法,在CityPersons数据集上达到了11.88%MR,比基准模型Adapted Faster RCNN在小尺度行人检测上有较大提升,相比于传统检测框架能更好的检测不同尺度的行人。该技术可以广泛应用在智能视频监控,车辆辅助驾驶等领域中。

With vigorous development of unmanned driving technology,pedestrian detection has become a major difficulty,and a hot research issue.To solve the problem of being ineffective of traditional pedestrian detection frameworks(One-stage and Two-stage,etc.)for small-scale pedestrian,the paper tries to improve pedestrian recognition accuracy of different scales in video sequence with new strategy based on FPN network.Firstly,the algorithm carries on multi-scale feature fusion based on ResNet50 feature extractration with FPN.At the same time,generate recommendation regions with RPN.Finally,carry on classification and regression for RPN generated recommendation regions with Fast RCNN,obtainfinal results after non-maximum suppression processing.Experimental results show pedestrian detection algorithm based on FPN achieves 11.88%MR on CityPersons data set,which is much better than small-scale pedestrian detection with benchmark model Adapted Faster RCNN,and can detect pedestrians of different scales better than traditional detection framework.The technology can be widely applied in intelligent video surveillance,vehicle assisted driving and other fields.