基于深度学习与多级匹配机制的港区人员轨迹提取

Port Staff Trajectory Extraction Based on Deep Learning and Multi-level Matching Mechanism

针对港口环境空间布局复杂,集装箱堆场、起重机械、装卸运输设备等复杂背景干扰下港区工作人员难以被准确跟踪的问题,本文面向港口监控视频提出一种基于Faster-RCNN(Faster Region Convolutional Neural Networks)检测算法和改进Deep SORT(Deep Simple Online and Realtime Tracking)跟踪算法的港区工作人员轨迹提取框架(FRIMDS)。本框架加入自适应高斯降噪和直方图均衡化算法,融合图像增强技术和行人重识别网络(Person Re-identification,ReID)提取港航图像特征信息,以提高港区工作人员轨迹提取的快速性和准确度。通过前置特征提取网络、候选区域建议网络、感兴趣区域池化和全连接层联合输出港区工作人员图像序列检测结果,采用级联匹配和匈牙利算法匹配港区工作人员位置信息,最后利用卡尔曼滤波预测得到港区工作人员运动轨迹。结果显示,本文所提方法在各典型港口场景中面对不同光照变化、低能见度、阴影干扰等挑战均表现出良好的性能,E_(IDF1)、E_(IDR)、E_(RCLL)、E_(MOTA)指标平均值分别为98%、97%、97%、95%。结论表明,本文提出的FRIMDS框架具有一定的精确性和稳定性,可为自动化码头安全监管提供技术支撑。

Due to the complex spatial layout of the port environment,the difficulty of accurate tracking of port staff exists under the interference of complex backgrounds such as container yards,lifting machinery,loading,unloading,and transportation equipment.This study proposes a trajectory extraction framework based on a Faster-RCNN detection algorithm and an improved Deep SORT tracking algorithm for port surveillance video.In this framework,an adaptive Gaussian noise reduction and histogram equalization algorithm were added,and the image enhancement technology and Person Re-identification network were integrated to extract the feature information of port images,to improve the rapidity and accuracy of the track extraction of port staff.The detection results of the port staff image sequence were output through the pre-feature extraction network,the candidate region suggestion network,the pool of interest area,and the full connection layer.The location information of port staff was matched by cascade matching and the Hungarian algorithm.Finally,the motion trajectory of port staff was predicted by the Kalman filter.The results show that the proposed method has good performance in the face of challenges such as different light changes,low visibility,and shadow interference in each typical port scene.The average values of EI_(DF1),E_(IDR),E_(RCLL),and E_(MOTA)are 98%,97%,97%,and 95%,respectively.The conclusion shows that the FRIMDS framework proposed in this study has certain accuracy and stability,and can provide technical support for the safety supervision of automated terminals.