基于深度强化学习的行人检测研究

Research on Deep Reinforcement Learning-based Pedestrian Detection

针对目前深度强化学习(Deep Reinforcement Learning,DRL)在目标检测中智能体初始化训练窗口固定单一、多目标和小目标图像错检率、漏检率高的问题,提出一种结合YOLOv5s和DQN算法的行人检测方法。该方法能够通过YOLOv5s搜索到含有目标的数量和区域,将回归的初步包围框设定为智能体的初始化窗口,提升尺度适应性。改进传统强化学习模型的奖励函数,使奖惩反馈更精准,提高模型检测精度和速度。与现有的基于深度学习、深度强化学习的目标检测模型对比实验,实测结果表明所提出的行人检测方法能够有效地提高检测精确度。

To address the current problems of fixed single initialization training window of intelligences in target detection by deep reinforcement learning(DRL),high error detection rate and leakage rate of multi-target and small target images,a pedestrian detection method combining YOLOv5s and DQN algorithm is proposed in this paper.The method is able to search the number and area containing targets by YOLOv5s,set the initial enclosing frame of regression as the initialization window of the intelligent body,and improve the scale adaptation.The reward function of the traditional reinforcement learning model is improved to make the reward and punishment feedback more accurate and improve the model detection accuracy and speed.Comparison experiments with existing target detection models based on deep learning and deep reinforcement learning are conducted to obtain empirical results showing that the proposed pedestrian detection method can effectively improve the detection accuracy.