激光与光电子学进展, 2017, 54 (8): 081003, 网络出版: 2017-08-02
基于加速区域卷积神经网络的夜间行人检测研究 下载: 769次
Nighttime Pedestrian Detection Based on Faster Region Convolution Neural Network
图像处理 红外图像 行人检测 加速区域卷积神经网络 区域建议网络 image processing infrared image pedestrian detection faster regional convolution neural network region proposal network
摘要
行人检测是机器人和无人车夜间工作应用中的重要任务之一, 采用加速区域卷积神经网络框架实现夜间红外图像中的行人检测, 用区域建议网络生成候选区域, 无需单独从图像中生成候选区域。区域建议网络和用于分类以及位置精修的卷积网络中, 采用卷积层参数共享机制, 使得该框架具有端到端的优点, 因此无需手动选取目标特征, 实现了从输入图像直接到行人检测的功能。实验结果表明, 与使用传统方法和快速区域卷积神经网络相比, 使用加速区域卷积网络框架对红外图像进行行人检测的准确率从68.2%和73.4%提高到了90.9%, 检测时间从3.6 s/frame和2.3 s/frame缩短到了0.04 s/frame, 达到了实际应用中的实时性要求。
Abstract
Pedestrian detection is one of the most important tasks of robots and unmanned vehicles at nighttime. Faster region convolution neural network framework is used to realize the pedestrian detection of infrared image at nighttime. This framework uses region proposal network to generate region proposals. Therefore, it is unnecessary to generate region proposals separately from the image. The parameter sharing mechanism is adopted in the convolutional layers in region proposal network and convolutional network for classification and bounding box regression, which makes the framework an end-to-end advantage. Thus, the pedestrian detection can be implemented from the input image to the detection result directly and it is unnecessary to manually select the features of the target. Experimental results show that the proposed method increases the recognition accuracy from 68.2% and 73.4% to 90.9% and shortens the recognition time from 3.6 s/frame and 2.3 s/frame to 0.04 s/frame compared with the traditional method and fast region convolution neural network, respectively, which reaches the required real-time level in practical applications.
叶国林, 孙韶媛, 高凯珺, 赵海涛. 基于加速区域卷积神经网络的夜间行人检测研究[J]. 激光与光电子学进展, 2017, 54(8): 081003. Ye Guolin, Sun Shaoyuan, Gao Kaijun, Zhao Haitao. Nighttime Pedestrian Detection Based on Faster Region Convolution Neural Network[J]. Laser & Optoelectronics Progress, 2017, 54(8): 081003.