联合语义的深度学习行人检测

视频行人检测是计算机视觉的一个重要应用，本文利用深度学习检测近似垂直视角的行人，但若单纯检测行人，易受与行人语义相关的行人附属属性（如背包和帽子）的干扰，容易造成误检.本文提出一种基于更快区域卷积神经网络的联合语义行人检测方法：首先调整网络模型，增强对小目标的辨别力，使其可以有效的检测行人和行人的语义属性；然后利用空间关系建立行人及其语义属性的关联，合并行人与其语义信息，并对候选行人目标进行自适应得分调整，结合行人语义属性判断候选行人目标.大量的实验表明，本文的方法精度高，速度快，具有实用价值，且检出的行人与其语义属性还可用于后续的人数统计和行人行为分析.

Deep Learning-Based Pedestrian Detection Combined with Semantics

Pedestrian detection is an important application of computer vision. However, it mostly uses the methods of low-level features. Deep learning, by combining the low-level features of pedestrians, can get more abstract representation of high-level features which makes the detection more robust. In this study, we propose a Faster Region-based Convolutional Neural Networks (RCNN)-based pedestrian detection method in which semantics is jointly considered. Firstly, we modify and fine-tune the Faster RCNN for fitting in the pedestrian dataset and for making it more capable of detecting small objects. Secondly, we establish connections between the pedestrian and its semantic attributes by spatial relationship, then fuse the pedestrian and its semantic attributes, and meanwhile adaptively adjust the confidence of the target pedestrian. The adaptive adjustment strategy, based on the connections between the pedestrian and its semantic attributes, realizes the fusion of the individual information. Extensive experiments and comparison show that the proposed approach in this study is of high accuracy, acceptable speed, and practical value. What is more, the semantic attributes can be used to count people or analyze the pedestrian''s behavior.