Pedestrian Detection Model Based On UAV Thermal Infrared Remote Sensing

Thermal infrared(TIR)is insensitive to the variation of light,it can work effectively in low light or no light conditions.At the same time,visible light has no penetrating power in fog and haze,but thermal infrared has the imaging ability under rain,snow,haze and other weather conditions.Using thermal infrared imaging data for target detection tasks can make up for the shortcomings of visible data.Therefore,developing an object detection model combing the advantages of thermal infrared remote sensing and the flexible maneuverability of UAV system for detecting ground people makes it performing a crucial role in fields,e.g.search,rescue,and security tasks under variation conditions.However,using UAV thermal infrared remote sensing data for ground personnel detection has the following problems:(1)Ground people occupies fewer pixels,which makes the general object detection model unable to accurately identify small targets from the perspective of UAV;(2)Small temperature difference between the target and the background,and unobvious texture differences result in the target gradually integrating with the background,and the detector cannot accurately detect the target;(3)The image obtained by the sensor in different flying attitudes and different sweep modes is quite different;(4)At present,the scientific community has not released relevant thermal infrared remote sensing data sets for the above problems.Therefore,the main object of this research include,(1)Establishing a new UAV-TIR remote sensing object detection dataset: UAVIRDet3500.First,DJI M600 Pro and FLIR Vue Pro R are integrated into a reliable UAVTIR remote sensing system.Based on the UAV-TIR system,2270 scenes TIR images were collected through many fields flying experiments.Meanwhile,two handheld TIR cameras were used to simulate the UAV flying experiments in Campus Qingshuihe,and,thus,collected 1230 scenes TIR images.Finally,the UAV-IRDet3500 dataset was established through professional visual interpretation.(2)Building a UAV-IRDetNet(UAV thermal Infrared remote sensing Detection Network)model for UAV thermal infrared remote sensing detection.UAV-IRDetNet is a one-stage target detection model based on deep neural network and is excellent in the detection of thermal infrared small targets.In modular testing,three feature extraction skeleton networks,i.e.ResNet,SE-ResNet,and ShuffleNet,were compared in this study.The results show that SE-ResNet performs better in the construction of the subsequent feature pyramid network,and ShuffleNet has the lowest mAP and Recall,thus,it is suitable for deploy on a mobile platform.We constructed 4 different feature pyramids and proposed a method for constructing a feature pyramid suitable for ground personnel detection.We found that predicting more feature pyramid layers does not improve the detection performance,and the construction of the feature pyramid should be more tended towards the bottom for people detecting tasks using the UAV-TIR system.In terms of the confusing problem between the target and background under nadir view,and large variance variations between targets under different observation attitudes,the hard case mining loss function is constructed and added in the detection model.The result indicates that the hard case mining improves the model's Recall and mAP.(3)Testing and comparative analyzing the UAV-IRDetNet model.The UAP-IRDetNet model has a mAP of 0.950 and a Recall of 0.991.The comparison result shows that the traditional general target detection algorithm fails to meet the assumption of small variance between targets and large variance between targets and background,thus,is not suitbal for UAV-TIR target detection task;When comparing with the general target detection algorithm based on deep learning,the result shows that the UAV-IRDetNet proposed in this paper is significantly improved,when compared to the two-stage general target detection model(Faster-RCNN)and the single-stage general target detection model(RetinaNet);The UAV-IRDetNet mentioned in this study has obvious advantages over infrared remote sensing target detection algorithms I-MRF and FKRWIn this paper,the UAV-IRDet3500 UAV thermal infrared remote sensing target detection dataset is established,and the UAV-IRDetNet UAV remote sensing target detection model is developed based on the characteristics of the data.The shortcomings of the traditional general target detection algorithm in the UAV thermal infrared remote sensing task are discussed.Lateral comparison is made between the general target detection algorithm based on deep learning and the dedicated infrared remote sensing target detection algorithm.This further taps the application potential of UAV thermal infrared remote sensing.