基于深层特征注意力提取的行人检测算法
Pedestrian Detection Algorithm Based on Deep Feature Attention Extraction
辛凯 1赵艳磊2
作者信息
- 1. 江苏大学计算机与通信工程学院 镇江 212001
- 2. 江苏航空职业技术学院 镇江 212001
- 折叠
摘要
行人检测是计算机视觉的基本挑战之一,目的是快速并准确的定位行人.针对行人检测中大目标难以定位、小目标难以检测以及拥挤人群的问题,通过增大深层特征图的感受野和丰富深层特征信息,提出了一种基于深层特征注意力提取的行人检测算法.首先,在骨干网ResNet-50的基础上为其添加新的阶段,通过空洞卷积进行更深的特征提取,以捕获丰富的上下文特征.其次,通过使用注意力机制从通道维角度挖掘特征间的关联信息,对行人特征进行加强.所提出的行人检测算法虽然结构简单,但它在具有挑战性的行人检测基准CityPersons和Caltech数据集上,展现了具有竞争力的准确性和良好的速度.
Abstract
Pedestrian detection is one of the basic challenges of computer vision.The purpose is to locate pedestrians quickly and accurately.Aiming at the problems of locating large targets and detecting small targets in pedestrian detection and crowded crowds,a pedestrian detection algorithm based on deep feature attention extraction is proposed by enlarging the receptive field of deep feature map and enriching the deep feature information.Firstly,a new stage is added to the backbone network ResNet-50,and deeper feature extraction is carried out through empty convolution to capture rich context features.Secondly,the attention mecha-nism is used to mine the correlation information between features from the perspective of channel dimension to enhance pedestrian features.Although the proposed pedestrian detection algorithm has a simple structure,it shows competitive accuracy and good speed on the challenging pedestrian detection benchmark CityPersons and Caltech datasets.
关键词
行人检测/卷积神经网络/空洞卷积/注意力机制Key words
pedestrian detection/convolutional neural network/empty convolution/attention mechanism引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据