基于多任务对齐的密集行人检测算法研究

Research on dense pedestrian detection algorithm based on multi-task alignment

行人检测是深度学习目标检测领域的重要分支,但密集场景中存在严重遮挡问题,给行人检测带来巨大挑战。为了缓解该问题,在CenterNet多任务学习模型上提出目标检测和姿态关键点检测任务对齐方法,改进后的模型为Center_tood。首先提出分离模块:该模块将原始特征分离得到更加关注各个任务的特征;在此基础上提出任务对齐方法:通过设计对齐度量来约束损失,使模型在梯度上更大程度地向着多任务对齐的方向优化,同时利用一致性约束,使模型学习到不同任务之间的共性信息,从而对齐不同任务的特征。实验部分采用CrowdPose数据集训练和测试。本算法的目标检测AP值为74.3%,提高了11.5%;人体姿态关键点AP值为55.8%,提高了9.6%。实验结果验证了提出的多任务学习算法在密集场景行人检测上的有效性。

Pedestrian detection is an important branch of deep learning object detection field,but there are serious occlusion problems in dense scenes,which brings great challenges to pedestrian detection.To alleviate this problem,a task alignment method for target detection and attitude key point detection was proposed on the CenterNet multi-task learning model,and the improved model was Center_tood.Firstly,the separation module is proposed.This module separates the original features into the features that pay more attention to each task.On this basis,a task alignment method is proposed:the alignment measurement is designed to constrain the loss,so that the model can optimize towards the direction of multi-task alignment to a greater extent on the gradient.At the same time,the consistency constraint is used to make the model learn the common information between different tasks,so as to align the features of different tasks.In the experiment part,CrowdPose data set was used for training and testing.The AP value of the proposed algorithm is 74.3%,which increases by 11.5%.The key point AP value of human posture was 55.8%,which increased by 9.6%.Experimental results verify the effectiveness of the proposed multi-task learning algorithm in pedestrian detection in dense scenes.