【作者】 姜敏；

【导师】 王力；

【作者基本信息】 贵州大学 ， 电子与通信工程， 2020， 硕士

【摘要】 行人检测技术是通过摄像头获取交通场景或监控视频下的视频图像信息,实现行人目标的精准定位。传统的行人检测算法多依靠人工设计特征,易受环境因素、人为因素的影响而达不到理想效果,深度学习因其优秀的特征学习能力在目标检测方面取得很大突破,但是复杂场景下的行人检测极易受到人体姿态多样、穿着各异和光照变化等影响而具有很强的挑战性。本文在深度学习目标检测算法SSD(Single Shot Multibox Detector)的基础上进行改进,以提升行人目标检测的性能,具体工作如下:1.针对SSD网络的各卷积层之间关联性不强、送入检测层的特征分辨率低导致目标定位不准确等问题,提出一种基于特征融合与附加特征提取的网络,对SSD网络的基础网络部分进行改进,将浅层特征与深层特征进行融合,使得网络的上下文信息能被充分利用,增强网络对行人目标的定位能力,在特征图进入检测层之前加入一个残差学习单元进行附加特征提取,达到降低计算量且增加特征表达能力的效果。实验表明特征融合及加入残差学习单元能一定程度降低行人目标的漏检率与误检率。2.针对原始SSD算法中浅层特征信息提取不充分,检测的特征图语义信息不充足的情况,提出一种引入注意力机制的SSD行人检测算法,以增强对行人检测任务有关的信息而抑制无关信息,达到增强行人特征的目的,通过在原来的基础网络VGG-16上添加通道注意力模块得到SSD-SENet网络,添加通道与空间注意力模块得到SSD-CBAM网络。实验表明,引入注意力机制的SSD行人检测算法相比原来的算法,漏检率和误检率均有所降低。更多还原

【Abstract】 Pedestrian detection technology is to obtain the video image information under the traffic scene or monitoring video through the camera,so as to achieve the accurate positioning of pedestrian target.The traditional pedestrian detection algorithm mostly relies on the artificial design features,which is vulnerable to the influence of environmental factors and human factors and can not achieve the ideal effect.Because of its excellent feature learning ability,deep learning has made a great breakthrough in the aspect of object detection.However,pedestrian detection in complex scenes is extremely challenging due to the influence of diverse human posture,different clothing and changes in light.Based on SSD(Single Shot Multibox Detector),the deep learning target detection algorithm was improved in this paper to improve the performance of pedestrian target detection.The specific work is as follows:1.Aiming at the problems such as the weak correlation between the volume layers of SSD network,the low resolution of the features sent to the detection layer and the inaccuracy of target location,a network based on feature fusion and additional feature extraction is proposed.The basic network part of SSD network is improved,and the shallow features and deep features are fused,so that the context information of the network can be fully utilized and the positioning ability of the network to the pedestrian target can be enhanced.The residual blocks are added before the feature maps enters the detection layer,which can reduce the calculation and increase the ability of feature expression.The experiment shows that feature fusion and the addition of residual blocks can reduce the rate of missing and false detection of pedestrian detection to a certain extent.2.In view of the insufficient shallow feature information extraction and semantic information of feature map in the original SSD algorithm,an attention mechanism based SSD pedestrian detection algorithm is proposed to enhance the information related to pedestrian detection tasks and suppress the irrelevant information,so as to enhance the pedestrian characteristics.SSD-SENet network is obtained by adding channel attention module to the original basic network VGG-16 and add channel and spatial attention module to the original basic network VGG-16 called SSD-CBAM network.Experimental results show that the SSD pedestrian detection algorithm with attention mechanism introduced in this paper have lower miss rate and false rate than the original algorithm.更多还原