Research On Multispectral Pedestrian Detection Algorithm In All-time Traffic Scene

Pedestrian detection is a key research topic in the field of computer vision,and has a wide range of applications in the field of intelligent transportation and autonomous driving.In recent years,the multispectral pedestrian detection method combining the information of visible image and infrared image has obvious advantages in the all-time detection environment,which has become a current research hotspot.However,due to the complexity and change of traffic road scenes,multispectral pedestrian detection algorithm will still be affected by changes in environmental conditions such as illumination and temperature,which makes it difficult to meet the accuracy requirements of practical applications of automatic driving.This thesis focuses on the optimization of multispectral pedestrian feature expression and fusion methods,and the main research contents are as follows.(1)Aiming at the problem that multispectral aggregate channel feature cannot characterize infrared pedestrian targets information sufficiently and characterizes targets inconsistently under different environmental conditions in daytime and nighttime,multispectral aggregate channel with entropy-weighted histogram of intensity difference feature for pedestrian detection in daytime and nighttime is proposed.Using the multispectral aggregate channel feature algorithm as the basic framework,when extracting features,the algorithm focuses on analyzing the advantages of infrared images to describe pedestrian targets.When constructing the new entropy weighted histogram of intensity difference feature,the algorithm uses neighborhood pixel intensity difference estimation and regional information entropy analysis to improve the original histogram of gradients feature.Due to the change of environmental conditions,the multispectral features are different in the daytime and nighttime.The algorithm uses adaptive boosting classifier to train the day and night images in the dataset to obtain day and night classifiers.In the detection stage,the HSV spatial histogram information of the image is used to distinguish between day and night,so as to realize the detection of the input image by time.The simulation experiment results show that the log-average miss rate of the proposed method in the all-time test scene is 5.93% lower than that of the original multispectral aggregate channel feature algorithm,which improves the performance of the pedestrian detector.(2)Aiming at the problem that traditional multispectral pedestrian detection networks have poor performance in all-time traffic scenes,an asymptotic localization fitting network multispectral pedestrian detection algorithm with attention mechanism is proposed.Under the framework of asymptotic localization fitting network,this algorithm uses a two-way Res Net50 with a convolution module as backbone network to extract the visible and infrared multi-scale features of pedestrian targets,and optimizes the features by combining the channel and space attention mechanisms.The size of the default candidate frame determined by the K-means clustering method is used to obtain the prior information of pedestrians.In addition,the algorithm researches the optimal fusion timing of the visible and infrared feature layers in the detection network,and the influence of the different positions of the attention mechanism module in the network on the detection effect.The simulation experiment results show that the algorithm has a log-average miss rate of 11.87% in the all-time test scene,which has a good detection effect for pedestrian targets in a complex traffic environment.(3)Based on the multispectral pedestrian detection framework of asymptotic localization Fitting Network,the fusion method of visible and infrared features is researched.Based on the different sensitivity of visible feature and infrared feature to the illumination change in the alltime traffic environment,a fusion method with illumination weight is proposed to improve the adaptability of the detection framework to the illumination change.Based on the complementary relationship between visible feature information and infrared feature information,a fusion method which is controlled by feature information is proposed,so that visible and infrared features can use their own specific information as weight for weighted fusion to improve the ability of the fusion feature to characterize pedestrian targets.The simulation experiment results show that the algorithm using illumination weight fusion method reduces the logarithmic average miss rate of the original algorithm by 2.01% in the all-time test scene,and the algorithm using feature information control fusion method reduces the logaverage miss rate of the original algorithm by 3.28% in the all-time test scene.