跨场景的行人关联方法研究

This project is based on the forensic intelligent application of large-scale surveillance videos. Its goal is to do research on the cross-scene pedestrian correlation. For this purpose, this project is to study the task from three aspects: pedestrian data collection and generation, pedestrian feature representation and pedestrian similarity measure. First, in consideration of existing surveillance being limited scenes and lacking insufficient data, a cross-scene pedestrian data collection and generation method is proposed, breaking the insufficient data limitation. Second, there exist differences in viewpoint and background under the surveillance environment of different scenes and time. To handle it, a pedestrian parsing based feature representation is presented, improving the robustness of feature; Finally, a similarity measure of high efficiency is put forward to solve the problem of existing methods being high complexity and low performance, enhancing the real time and efficiency of computing the similarity. This project is expected to improve the accuracy and real time of cross-scene pedestrian correlation under surveillance environment, advance both academy research and real applications of intelligent video analysis technologies.

本项目面向公安视频侦查中海量监控视频的智能分析应用，旨在研究跨场景的行人关联问题。为此，本项目拟从行人数据采集与生成，行人特征表达和行人相似度度量三个方面进行研究。首先，针对现有监控数据场景单一、行人关联数据缺乏的问题，拟研究跨场景行人数据采集和生成方法，突破数据约束；其次，针对不同场景和时间下的监控环境存在视角和背景的差异，提出基于行人解译的特征表达算法，提升特征的鲁棒性；最后，针对现有行人相似度算法复杂度高且性能差的问题，提出一种高效的相似度计算方法，提升相似度计算的实时性和有效性。本项目的研究有望提高跨场景行人关联的准确率和实时性，并推动学术研究的发展和智能视频分析技术的实际应用。

本项目构建了一个跨场景行人关联平台。研究了跨场景行人关联问题，提出基于颜色转换的生成对抗网络和无监督图关联的行人再识别算法，生成新的行人数据，解决训练数据欠缺问题；提出基于注意力机制和级联式“分离-聚合”学习模块，学习鲁棒的行人高层语义信息；提出跨视角联合学习策略解决特征维数高、训练效率差等问题。收集、整理了多路行人视频，有效促进了跨场景行人关联领域的发展。