基于深度学习的行人属性识别

Pedestrian Attribute Recognition Based on Deep learning

针对视频图像中的行人分辨率低、不同场景中存在着光照明暗的变化、视角不同的变化、姿态各异变化以及被物体遮挡等问题,导致行人属性(性别、年龄、穿衣风格等)检测变得困难,提出基于caffe框架的深度卷积神经网络进行行人属性识别。利用行人数据库训练网络模型,所用模型可以输入任意大小的图像而不需进行缩放或裁剪。此网络可以同时完成对行人92个属性的识别,端到端的训练整个卷积神经网络,算法简洁高效,和现有的其他方法进行比较,即使在有限的数据支持下,依然能够取得不错的性能,在行人数据库RAP数据集上获得较高的准确率。

Pedestrian attribute（such as gender,age,dressing style ect.）recognition in surveillance scene is also a challenging prob-lem due to the low resolution,the change of illumination in the scene,the large pose variations arisen jfrom different angles of view,occlusions from environmental objects.We propose a deep learning convolutional neural network based on Caffe frame-work for pedestrian attribute recognition.In this work,we using a pedestrian database to train network,without scaling or war-ppping the images.This network can recognition pedestrian attribute at the same time ,end-to-end training the neural network,the algorithm is simple and efficient,and compare with other existing methods,even in the limited data support,we can still achieve a good performance in accuracy of RAP（Richly Annotated Pedestrian）dataset.