Research On Pedestrian Target Detection Algorithm Based On Infrared Image

In the field of vehicle assistance driving systems and security monitoring,the detection of pedestrian targets is crucial.Video image based on visible light have poor imaging quality or cannot be imaged in an environment with insufficient illumination,and a large amount of effective information is destroyed,making it difficult to effectively recognize pedestrian targets.Compared with visible light,infrared image does not depend on ambient lighting conditions,reflecting the difference in surface temperature of the target surface with different thermal radiance rates.It can also be effectively imaged in environments with poor visibility such as dark nights and rainy days,and has a wide range of applications,therefore pedestrian target detection algorithm based on infrared image has important application prospects;Pedestrian target action is diverse,randomness is strong,appearance mode is complex and variable,scale is different.Compared with visible light,infrared pedestrian image lacks color information and the texture information is less.The feature distinguishability is not high,which makes it difficult to design pedestrian detection algorithms.Therefore,the pedestrian detection algorithm based on infrared image still has research value and space for improvement.The method proposed in this paper is based on the deep learning network,and improves the Single Shot Multibox Detector(SSD)algorithm for the road pedestrian target.The main work is as follows:(1)We improved the feature extraction network.The expansion channel convolution method was used to aggregate the features of different receptive fields,enhance the feature extraction ability of convolutional networks.We built an I-SSD infrared pedestrian detection network with improved basic network,designed network connection methods and parameters,and improved the detection ability of network for pedestrian target.(2)The feature channel weight recalibration module was added to the network to learn the modeling of the relationship between different feature channels through the end-to-end training of the entire network.It could calculate the importance weights of different feature layers,and improve the feature layer which can play a more important role in pedestrian target detection.The clustering algorithm is used to analyze the pedestrian target boxes,and the network default box parameters are set according to the analysis result to improve the targeting of the pedestrian.The IS-SSD network is built to further improve the detection effect of the network.(3)The detection algorithm proposed in this paper is implemented on the experimental platform.Different algorithms were tested based on the self-built infrared image dataset to compare the performance of different algorithms.The precision and recall rate of the IS-SSD algorithm proposed in this paper are 3.43% and 4.27% higher than the original SSD algorithm.The experimental and test results show that the infrared image pedestrian detection algorithm proposed in this paper has high precision and recall rate,which could meet the needs of practical applications.