基于深度学习的行人属性多标签识别

Multi-Label Recognition of Pedestrian Attributes Based on Deep Learning

行人属性通常指的是行人的一些可被观察到的外部特征,如性别、年龄、服饰、携带品等。作为行人外部的软生物特征,行人属性对于行人检测和再识别是非常重要的,并且在智能视频监控场景和基于视频的商业智能应用中显示出巨大的潜力。在目前的行人属性多标签分类识别中,主要有基于手工设计特征的方法和基于深度学习的方法。然而,手工设计特征的方法难以应对复杂的真实视频监控场景,在实际应用中取得的效果并不是很理想。采用深度卷积网络模型,包含3个卷积层和2个全连接层,使用Sigmoid交叉熵损失函数,训练平台为Caffe深度学习框架,通过在包含19 000张行人图片的PETA数据集上对10种行人属性进行训练和测试,得到85.2%的平均识别精度。加入正样本比例指数因子改进损失函数后,平均识别精度达到89.2%,使网络性能有明显的提高。

Pedestrian attributes usually refer to some of the external characteristics of pedestrians that can be observed,such as gender,age,clothing type,carrying objects,etc. As soft biological features of pedestrians,pedestrian attributes are very important for pedestrian detection and re-identification,and show great potential in intelligent video surveillance scenarios and video based business intelligence applications. Among the current multi-label classification methods of pedestrian attributes,two of them are mainly employed,one is based on handcrafted features and the other is based on the deep learning methods. However,the methods of handcrafted features are difficult to deal with complex real video surveillance scenes, results obtained in practical applications are not ideal. In this paper we used a deep convolutional network model with three convolutional layers and two full-connected layers. Using the Sigmoid cross-entropy loss function,the training platform was the Caffe deep learning framework,the dataset used was PETA containing 19,000 pedestrian images. Ten kinds of pedestrian attributes were trained and tested,and an average recognition accuracy of 85. 2% was reached.After adding the positive sample proportional exponential factor to improve the loss function,the average recognition accuracy reached 89. 2%,which significantly improved the performance of the network.