Research On Pedestrian Detection In Infrared Image Based On Deep Learning

At present,pedestrian detection at night mainly uses visible light image,lidar,infrared image and so on.At night,due to the unsatisfactory lighting conditions,the imaging and environmental monitoring effects of detectors such as visible light cameras are poor,while infrared imaging technology is favored by people for its all-weather characteristics,which is based on thermal radiation imaging and is not affected by the light.In recent years,with the rapid development of deep learning technology,deep learning has become the mainstream method of many pattern recognition and object detection problems.The pedestrian detection technology based on deep learning infrared image has great market application potential.This article first analyzes the target,background,and noise characteristics of the infrared image.Target pre-processing of infrared pedestrian images.In terms of denoising,this article introduces some traditional denoising methods.The infrared pedestrian image is preprocessed.In the aspect of de-noising,this paper introduces some traditional de-noising methods,and proposes a wavelet adaptive threshold de-noising method combined with wavelet de-noising.The image is decomposed by wavelet,and the threshold of each layer is obtained by calculating the scale parameters,so as to de-noising the high-frequency part of different layers.In addition to filtering noise,it also retains better edge information,and the processing effect is better than other methods.In contrast enhancement,the adaptive histogram equalization algorithm with limited contrast has good results.In order to improve the contrast of the image,neither the user input nor the over enhancement of the image is needed,and the processing effect is better than other methods.This paper also introduces the basic knowledge of convolutional neural network,as well as the existing mainstream deep learning platforms and models.Through comparative analysis,it is found that SSD network has greater advantages.Therefore,an improved method of pedestrian detection in infrared image of SSD network is proposed.Because the SSD network uses VGG-16 for feature extraction and the network computation is large,it is replaced by Mobile Net V2 network with high accuracy and less parameters,so that the network can meet the real-time requirements while ensuring the accuracy.K-means algorithm is used to adjust the number of prior boxes and the ratio of length to width,so that the network is more suitable for pedestrian detection.The test results show that the improved network detection time is shortened and the accuracy is improved.The m AP can reach 91.73% on CVC-09 data set.This paper proposes SSD algorithm based on transfer learning to solve the problems of low image detail and low detection rate in infrared image pedestrian detection.The new Moblie Net V2(1.4)+SSD network parameters are initialized based on the model weights of the original Moblie Net V2(1.4)+SSD iteratively trained on the CVC-09 dataset Then the 6-layer network weights directly related to the output layer are re-learned to accelerate network convergence.After the OUS thermal infrared database test,m AP reached 94.6%.Experimental results show that the proposed method can effectively speed up the detection speed and improve the recognition accuracy.