Pedestrian Detection Method In Infrared Image

Pedestrian detection is an important branch in the field of object detection.It is widely applied in the field of intelligent security,vehicle assisted driving and intelligent transportation for social and economic value.In recent years,it has became a research hotspot in the fields of computer vision,pattern recognition,and machine learning.Paticularly,benefiting from the successful application of deep learning in computer vision,pedestrian detection technology has been developed rapidly.Infrared imaging can solve the failure of visible light imaging at night,therefore,by combining two imaging mode,an allweather pedestrian detection system can be constructed.However,compared with visible light imaging,infrared imaging has the disadvantages of single channel,low contrast,low signal-to-noise ratio,etc.,and there are fewer public datasets.These drawbacks make infrared image pedestrian detection algorithms more challenging to design,as well as the potential more trenmendous.This thesis focuses on the two problems in infrared image pedestrian detection,the false alarms in complex environments and the insufficient generalization capabilities of detection models.Applying Efficient Det,a lightweight one-stage object detection network,multi-task learning is introduced,two pedestrian detection models based on convolutional neural networks are designed to improve the detection performance for infrared image.The specific work content is summarized as follows.(1)For the false alarm problem in infrared image pedestrian detection,a multi-task pedestrian detection algorithm based on segmentation constraints is proposed.A lightweight semantic segmentation branch is embedded into an object detection network,the pedestrian area is estimated by segmentation branch of shared feature extraction network,and the pedestrian position is constrained by the scene semantics obtained from the segmentation task,which can reduce background false alarms.The receptive field of the model is extended by the atrous spatial pyramid pooling in the semantic segmentation branch,the convolutional block attention module is used for feature optimization,and the feature fusion with different levels is used to enhance the model's robustness to different scales targets.Finally,logical operations between semantic segmentation output and the detection result is uiltilized to filter out false alarms.Experimental results show that on the condition of the same recall rate,the accuracy rate is effectively improved.(2)For the problem of weak model generalization ability caused by the homogenization of the infrared image dataset,the transfer learning is introduced to the pedestrian detection algorithm based on supervised domain adaptation.Based on the object detection framework,the domain adaptation branch is embeded to align the feature between infrared image and visible light image,and so the visible light image dataset can be a training subset of the infrared image dataset.For the domain adaptation branch,firstly,the alignment between local feature and global feature is designed by the domain classifier combined with the gradient reversal layer;secondly,the consistency constraints between the local and the global output of domain classifier are carried out to ensure the consistency of the local and global feature domain classification results.Experimental results show that this method can effectively improve the performance of pedestrian detection on infrared images.