基于深度学习的视频行人目标检测

Video Pedestrian Detection Based on Deep Learning

摘要: 近年来，随着深度学习在计算机视觉领域的广泛应用，基于深度学习的视频运动目标检测受到广大学者的青睐。这种方法的基本原理是利用大量目标样本数据训练一个基于深度神经网络的分类器，然后通过分类器在线检测目标。由于深度神经网络能够通过多层表示的方式更加深刻的描述目标特征，基于深度学习的检测方法优点在于能够准确检测具有训练数据中目标特征的目标。针对视频运动目标检测这个特定的应用，这种方法的局限性在于没有利用目标运动信息，检测结果容易出现虚警目标。本文将GMM建模方法与深度神经网络相结合，充分利用目标外观特征和运动信息，以期获得更准确的检测结果。在2017年央企双创展实地采集的展台监控数据上进行了实验验证。结果表明，本文方法相比于不融合运动信息的检测方法，行人检测准确率提高3.8%。

Abstract: In recent years, with the extensive application of depth learning in the field of computer vision, video moving target detection based on deep learning has been favored by many scholars. The basic principle of this method is to train a classifier based on deep neural network using a large number of target sample data, and then detect the target online through the classifier. Since the depth neural network can describe the target features more deeply through multi-layer representation, the detection method based on deep learning has the advantage of being able to accurately detect targets with features from training data. For this particular application of moving target detection, the limitation of this method lies in that it does not use the target motion information and the detection results are prone to false alarm targets. In this paper, GMM method is combined with deep neural network to make full use of the target appearance and motion information in order to obtain more accurate detection results. The experimental verification was carried out on the monitoring data collected from the exhibition booth of the State-owned Enterprise Exhibition in 2017. The results show that compared with the detection method without fusion of motion information, the accuracy of pedestrian detection is improved by 3.8%.