【作者】 罗来全； 李升波； 高洪波； 刘征宇； 成波；

【Author】 LUO Lai-quan;LI Sheng-bo;GAO Hong-bo;LIU Zheng-yu;CHENG Bo;State Key Laboratory of Automotive Safety and Energy,Tsinghua University;

【机构】 清华大学汽车安全与节能国家重点实验室；

【摘要】 为了快速准确地在复杂环境中进行目标检测,本文提出一种基于深度强化学习算法的图像检测方法。将深度学习感知能力与强化学习决策能力相结合,利用深度学习预训练网络提取特征,利用强化学习将检测过程视作马尔可夫决策过程问题,将当前可视区域视为状态,将边界框变换作为动作,基于重叠度值变化设计相应的奖励函数。为检验提出方法的有效性,通过Pascal VOC数据集进行验证,实验结果表明,在对图像进行少于10个区域的分析情况下,能快速检测出图像中车辆的目标,且精度达到58.4%,每张图片平均检测时间为6.04s。与深度学习目标检测算法对比,该方法极大地减少了候选区域数量,所需计算量较低,在保证准确度的前提下,具备检测速度快、准确高的特点,这为复杂环境中地面无人平台的目标检测提供了新型的方法。更多还原

【Abstract】 In order to accelerate the process of object localization in a complex environment,a deep reinforcement learning based object detection method is proposed.Combining the perception ability of deep learning with the decision-making ability of reinforcement learning.In the deep learning part,the pre-training network is used to extract features.In the reinforcement learning part,the object detection process is regarded as a Markov decision process.The current visual regions are regarded as states,the boundary box transformations are used as actions,the reward function is designed based on the change of the intersection-over-union with the truth box.In order to verify the effectiveness of the proposed method,Pascal VOC dataset is used for testing.The experimental results show that the vehicle target in the image can be detected with an accuracy of 58.4% and an average detection time of 6.04 s per image when the image is analyzed in less than10 regions.Compared to object proposal algorithms,this method reduces the candidate regions and requires less computation.It reduces the detection time while guarantee the detection accuracy,which proves that this method is very efficient.It provides a new idea and method for target detection of UGV(Unmanned Ground Vehicle) in complex environments.更多还原