| In recent years,as the country attaches great importance to autonomous driving technology,pedestrian detection system as the core link of autonomous driving has also received widespread attention.It can be applied to intelligent fields such as traffic detection,video surveillance,and intelligent robots.Although the current pedestrian detection algorithm based on visible images has made great progress,the performance of the detector under severe conditions such as low light or night will drop sharply.The use of visible and far-infrared image fusion can effectively solve the problem of all-weather pedestrian detection.However,the current pedestrian detection method based on the fusion of visible and far-infrared image features(referred to as multispectral pedestrian detection)has problems such as high computational complexity,inaccurate detection boxes positioning,weak expression of pedestrian pattern features,and low detector operating efficiency.To solve the above problems,this paper mainly studies high-performance real-time multispectral pedestrian detection in all-weather scenarios,using the visible and infrared image feature fusion algorithm based on deep learning as the entry point,which has a high rate of missed detection of small-scale targets and high model calculation complexity.Issues such as feature redundancy and blind fusion of multi-modal features have been studied in depth.The research work of this paper is as follows:(1)Aiming at the problems of low detection accuracy and high computational complexity of small-scale targets,a multispectral pedestrian detection algorithm based on improved Retina Net is proposed.This method first uses the dual-stream Retina Net as the backbone network of the multispectral pedestrian detection algorithm,and then adds a feature pyramid module to improve the spatial and semantic resolution of features.Aiming at the problem of too many low-quality detection boxes,an improved loss function is proposed to remove a large number of low-quality detection boxes to improve the quality of the overall detection boxes.The experiment results on the KAIST public data set show that the average miss rate of this method is 8% lower than that of the Faster R-CNN method.(2)Aiming at the problem that the output feature map of convolutional neural network contains a lot of noise information,a multispectral pedestrian detection algorithm based on improved spatial and channel attention is proposed.This method uses the improved space and channel attention modules to solve the problem of feature information redundancy and noise aliasing from the two dimensions of space and channel respectively.The experiment results on the two public data sets of KAIST and CVC-14 show that the average miss rate of this method is 19.44% and25.6%,respectively.Compared with the method without the attention mechanism,the miss rate is reduced by 2 percentage points.(3)Aiming at the problem that the multispectral pedestrian detection model takes too long to detect,a multispectral pedestrian detection algorithm based on the anchor-free mechanism is proposed.This method introduces the anchor-free mechanism of the Center Net model for the first time,which greatly reduces the computational complexity of the model.The detection speed(FPS)of the model is twice that of the Faster R-CNN detector.Aiming at the problem of blind fusion of multi-modal features,this method introduces a differential feature perception fusion module in the feature fusion network,and uses cross-modal complementary differential information to enhance the expression ability of original features and further improve the detection performance of the model.The experiment results on the three public data sets of KAIST,CVC-14 and FLIR show that the average miss rate of this method is 14.51%,25.54% and 28.43%,and the operating speed reaches20 FPS,which is compared with the anchor-base method of the Faster R-CNN algorithm has reduced the miss rate by 11 percentage points and doubled the detection speed. |
