基于深度学习的行人属性识别及应用

Pedestrian Attribute Recognition and Application Based on Deep Learning

为了提高行人属性识别的准确率,提出基于多尺度注意力网络的行人属性识别算法,并对其进行了改进。将行人属性分为上身、上身附属、下身、下身附属、脚部、朝向、性别、年龄、携带物9个类型。在进行初步的二分类属性预测后,实行进一步的属性筛选分类,避免互斥属性同时出现,从而提高属性识别结果的准确性和合理性。此外,为了有利于行人属性识别算法的应用,基于模块化的设计理念,按照原始图像和目标检测预测结果中的行人类型和位置信息获取行人图像块信息进行属性识别,提出目标检测与行人属性识别一体化的方法。

In order to improve the accuracy of pedestrian attribute recognition,a pedestrian attribute recognition algorithm based on multi-scale attention networks is proposed and improved.Divide pedestrian attributes into 9 types:upper body,upper body attachment,lower body,lower body attachment,feet,orientation,gender,age,and carrying items.After preliminary binary attribute prediction,further attribute screening and classification are implemented to avoid the simultaneous occurrence of mutually exclusive attributes,thereby improving the accuracy and rationality of attribute recognition results.In order to facilitate the application of pedestrian attribute recognition algorithms,and based on a modular design concept,pedestrian image block information is obtained from the original image and the pedestrian type and position information in the target detection prediction results for attribute recognition.An integrated method of target detection and pedestrian attribute recognition is proposed.