| Pedestrian recognition is an important research direction in computer vision and is widely used in multiple fields such as intelligent security,assisted driving,and military reconnaissance.However,outdoor pedestrian images are often affected by harsh environments such as low lighting,rain,and fog,and single-modality pedestrian recognition can easily result in false positives and false negatives,which affects subsequent recognition performance.In response to the above issues,this article adopts the fusion of infrared and visible light modes,and constructs the MSKE-YOLO pedestrian recognition algorithm based on the construction of a multimodal pedestrian dataset in harsh environments to improve the accuracy of pedestrian recognition in harsh environments.The main research contents of this paper are as follows:(1)Preprocessing and construction of pedestrian database in harsh environmentsTo address issues such as poor image quality,low data volume,and insufficient sample diversity in the pedestrian database under harsh environmental conditions,this study first employs an improved Retinex algorithm and an improved adaptive enhancement algorithm to enhance the image quality.Subsequently,high-quality images obtained are augmented using traditional methods and a DCGAN model to enrich the sample diversity.Expand and solve the problem of sample imbalance,and finally mark and divide the constructed pedestrian database in harsh environments for the subsequent application of image fusion and pedestrian recognition models.(2)Research on multimodal image fusion algorithm based on improved GAN networkIn view of the problems of gradient disappearance,slow training speed and easy to ignore image detail information in traditional image fusion methods,this paper designs SDGAN network model to fuse infrared and visible light images.Firstly,an attention mechanism module is introduced in the generator to enhance channel feature attention and a dense connection is added to solve the problem of gradient vanishing.Secondly,a dual-discriminator structure is used with spectral normalization to improve model training stability and accelerate model training.Finally,the design Detail loss and object edge enhancement loss to sharpen object edges.Through qualitative and quantitative analysis via simulation experiments with five classic fusion algorithms,the results demonstrate that the SDGAN model can obtain fusion images with clear edges and rich texture details,and can be applied to subsequent pedestrian recognition and related fields.(3)Research on multimodal pedestrian recognition algorithm based on improved YOLOv5In response to the problem of low accuracy and poor robustness of existing pedestrian recognition models in harsh environments,this paper proposes the MSKE-YOLO pedestrian recognition algorithm.Firstly,the Mosaic-9 data augmentation method is used to improve the robustness of the model.Secondly,the attention mechanism is introduced to enhance the model’s ability to extract features of pedestrian targets in harsh environments.Finally,K-means++clustering is employed to accelerate model convergence,and the bounding box regression loss function is improved to enhance model accuracy.Through comparisons with single-modal and multi-modal methods,four classic algorithms,and ablation experiments,the results show that the MSKE-YOLO algorithm achieves an accuracy of 93.1% and a real-time processing speed of 43 FPS.The improved algorithm is also verified on a target detection and tracking control system platform to demonstrate its practicality,and can be effectively applied to multi-modal pedestrian recognition and related fields in harsh environments. |
