基于时序信息和注意力机制的视频目标检测  

Video Object Detection with Temporal Information and Attention Mechanism

作　　者：蔡强[1,4,5] 李韩玉 李楠 刘新亮[3,4] CAI Qiang;LI Han-yu;LI Nan;LIU Xin-liang(School of Computer Science and Engineering,Beijing Technology and Business University,Beijing 100048,China;School of Artificial Intelligence,Beijing Technology and Business University,Beijing 100048,China;School of E-commerce and Logistics,Beijing Technology and Business University,Beijing 100048,China;National Engineering Laboratory for Agri-Product Quality Traceability,Beijing 100048,China;Beijing Key Laboratory of Big Data Technology for Food Safety,Beijing 100048,China)

机构地区：[1]北京工商大学计算机学院,北京100048 [2]北京工商大学人工智能学院,北京100048 [3]北京工商大学电商与物流学院,北京100048 [4]农产品质量安全追溯技术及应用国家工程实验室,北京100048 [5]食品安全大数据技术北京市重点实验室,北京100048

出　　处：《计算机仿真》2021年第12期380-385,共6页Computer Simulation

基　　金：北京市社会科学基金项目(19GLB036);国家社会科学基金项目(18BGL202);北京市自然科学基金(L191009);北京市教委科研团队建设项目(PXM2019_014213_000007);北京市科技计划(Z191100008619007)。

摘　　要：目前,基于深度学习的图像目标检测算法已经趋于成熟,但其无法利用视频中独特的时序信息,会使检测精度产生大幅度的下降。为了更好的进行视频目标检测,应当充分挖掘视频图像之间的联系,利用视频中的时序信息。因此提出了基于时序信息和注意力机制的视频目标检测算法(TIAM)。算法中加入了运动历史图像,表征视频中的时序信息,并为模型提供目标的运动信息;结合注意力机制,使模型更加关注目标区域,提高了区域特征的代表性。在大规模数据集Image Net VID上进行实验,验证了算法的有效性,平均精度均值达到了先进水平。Still-Image object detection based on deep convolutional neural networks has achieved good perform-ance at present. However, it cannot make use of the unique temporal information in a video, which will greatly reducethe detection accuracy. In order to better detect video objects, it is necessary to fully study the special temporal infor-mation in the video and explore the relationship between video images. Therefore, a video object detection algorithmbased on temporal information and attention mechanism(TIAM) is proposed. In this algorithm, the motion historyimage is added to represent the temporal information in the video, which provides the motion direction of the objectfor the model. By introducing the attention mechanism, the model pays more attention to the object area and improvesthe representation of the regional features. Experiments were carried out on the large-scale dataset Image Net VID toverify the effectiveness of the algorithm. And the mean average precision has reached an advanced level.

关 键 词：计算机视觉 视频目标检测 运动历史图像 特征融合 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]