基于改进FairMOT的铁路周界入侵检测方法

Railway Perimeter Intrusion Detection Method Based on Improved FairMOT

针对铁路综合监控视频中不同远近行人成像面积差异较大、自然环境变化产生干扰等因素造成的检测难题,提出一种改进FairMOT框架的周界入侵检测方法。首先,针对监控视频中不同远近的行人,通过在FairMOT框架中引入感受野模块,丰富不同成像大小行人检测所需的感受野,以更好地提取不同尺度特征信息;其次,针对夜晚时段方法检测性能较低的问题,在编码解码网络后融合空间注意力模块,强化夜间前景行人关键特征,同时优化目标跟踪和判断流程,实现稳定检测;然后,针对缺乏大量学习样本的问题,使用行人检测跟踪数据集与铁路真实数据集混合增强训练,提高方法在全天候检测中的泛化性和鲁棒性;最后,在MOT17数据集和铁路真实数据集上,对改进FairMOT检测方法与CenterTrack,Bytetrack等方法进行对比试验。结果表明:提出的改进FairMOT检测方法在白天和夜晚对不同大小目标检测中,均取得了最高的准确率和召回率调和均值,检测性能最好;方法检测速率为25.2帧·s^(-1),能够满足实时检测要求。改进的FairMOT检测方法可以更有效地应用于实际铁路周界入侵检测场景。

Aiming at the detection problems caused by the large difference in the imaging area of pedestrians at different distances and the interference caused by changes in the natural environment in comprehensive railway video surveillance,a perimeter intrusion detection method based on the improved FairMOT framework is proposed.First of all,for pedestrians at different distances in the video,the Receptive Field Block(RFB)module is introduced into the framework of FairMOT to enrich the receptive fields required for pedestrian detection of different imaging sizes,so as to better extract the feature information of different scales.Secondly,to solve the problem of low detection performance of the method at night,the spatial attention module is integrated after the encoding and decoding network to strengthen the key features of pedestrians in the foreground at night while optimizing the target tracking and judgment process to obtain stable detection at the same time.Then,for the lack of a large number of learning samples,this research uses the pedestrian detection tracking data set and the real railway data to enhance the training so as to improve the generalization and robustness of the method in all-weather detection.Finally,on the MOT17 dataset and the real railway dataset,the improved FairMOT detection method is compared with CenterTrack,Bytetrack and other methods.The results show that the proposed improved Fair MOT detection method has achieved the highest average score and the best detection performance in the detection of objects of different sizes during the day or night.The detection rate of the method is 25.2 fps,which can meet the real-time detection requirements.The improved FairMOT detection method can be more effectively applied to the actual railway perimeter intrusion detection scenarios.