Study On Pedestrian Detection Under Haze Weather

With the rapid development of artificial intelligence technology,target detection technology has been widely used in various fields.As a branch of target detection,pedestrian detection technology has also made a breakthrough,becoming one of the necessary key technologies in smart city,intelligent monitoring,intelligent home,autonomous driving and other fields.However,with the frequent haze weather of haze,complex weather conditions have an impact on pedestrian detection technology.It are still certain challenges to eliminate the impact of haze,and ensure that pedestrian detection technology has high accuracy in the background of haze weather.Therefore,this thesis conducts the following research on pedestrian detection tasks in haze weather:(1)A sky segmentation demogging scheme based on the dark channel demogging algorithm is proposed.According to the boundary effect of constant constant,OTSU sky segmentation algorithm divides the sky into the sky area,and introduce gradient-oriented filtering to optimize the transmission,so as to improve the clarity of the image defog.We proved the effectiveness of the algorithm using the information entropy,the mean and the standard difference as the image quality and definition.To provide help for the improvement of the pedestrian detection accuracy under the subsequent haze.(2)In terms of pedestrian detection,a SSD-based feature fusion pedestrian detection algorithm is proposed,using VGG-16 as the basic network for pedestrian detection,and constructing a pyramid SSD network feature layer for feature extraction on the basic network.In view of the detection of small targets and leak detection during SSD network pedestrian detection,three void convolution was added to the VGG-16 feature layer to expand the sensory field of the feature layer,increase the scope of network feature extraction,and improve the feature extraction ability of the shallow feature map on the target.In view of the contextual connection between characteristic layers of SSD network is not strong and it is difficult to integrate different scales,the integration of different characteristic layers is realized by adding characteristic corresponding elements,strengthens the judgment of the position information and semantic information of pedestrians,improves the detection accuracy of small target pedestrians,and effectively reduces the leakage detection rate.The Focal Loss loss function is introduced to balance the positive and negative samples during training,focusing on negative samples with rich available information,thus promoting network convergence while becoming more stable.This thesis studies the detection capability of the existing SSD pedestrian detection algorithm on the open dataset,proposes the improvement scheme,and trains and tests the model with the open dataset.Through experimental verification,the improved SSD algorithm is available and combines the improved dark channel algorithm with the SSD pedestrian detection algorithm.The experimental results show that the SSD-based feature fusion detection network is better than the accuracy before defogging.It shows that image ogging can improve the performance of pedestrian detection network and has some reference value for the target detection research in subsequent complex weather.