Research On Pedestrian Detection Technology Based On Multi-feature Fusion

Pedestrian detection is an important research direction in computer vision.In the field of scientific research,it provides technical support for research on target tracking,pedestrian reidentification,gait recognition.In production and life,it is in assisted driving,the intelligence of the disaster area.rescue and intelligent monitoring have a wide range of applications.With the rise of deep learning,the pedestrian detection technology based on deep learning has been greatly improved compared with the traditional methods,but the current pedestrian detection research still has many difficulties.The background of pedestrian is often complicated.When extracting features from input images,how to improve the feature extraction ability of the input image is a difficult problem to solve.In addition,the occlusion problem in dense scenes has a certain impact on the accuracy of detection.In this paper,the two problems mentioned above are studied.Based on the multi-feature fusion based pedestrian detection feature extraction network,two new pedestrian detection methods are constructed.The first method is to combine the Resden network proposed in this paper with the traditional edge feature extraction method to improve the detection accuracy of pedestrian detection.The second method is to add the IOU-loss in the pedestrian detection algorithm,and combine the Squeezeand-Excitation blocks and the Resden block in different ways to solve the problem of pedestrian occlusion.Firstly,the clustering algorithm is used to obtain the anchor used in the pedestrian detection algorithm in the data set.In order to improve the feature extraction ability of the network for the input image,the residual block and the dense block are combined into a Resden block,feature extraction of input images using a Resden network of Resden block.At the same time,for the multi-scale problem in pedestrian detection,this paper adds the feature pyramid structure to the Resden network,and proposes a pedestrian detection feature extraction network based on multi-feature fusion to achieve multi-scale prediction of pedestrians.The effectiveness and rationality of the pedestrian detection feature extraction network based on multi-feature fusion is proved by a large number of contrast experiments.Based on this,a multi-feature fusion pedestrian detection method combining traditional features and a multi-feature fusion pedestrian for densely occluded pedestrians are proposed.For the multi-feature fusion pedestrian detection method combined with traditional features,this paper studies the specific features of the input image extraction by different depth convolution layers,and then concluded that the low convolutional layer output feature map mainly contains the edge information of the input image.Therefore,in order to improve the feature extraction ability of the network from the low convolution layer,the traditional Canny edge detection operator is used to detect the edge of the input,and the obtained fine edge feature is combined with the feature of the low convolution layer output.The final experimental results prove that the multi-feature fusion pedestrian detection method combined with traditional features proposed in this paper can effectively improve the accuracy of pedestrian detection.For the multi-feature fusion pedestrian detection method for densely occluded pedestrians,this paper analyzes the reasons why the occlusion between the descendants of dense scenes brings difficulties to pedestrian detection,and then proposes the Intersection Over Union(IOU)loss function.By adding the IOU-loss to the commonly used loss function,the detector's tightness to the pedestrian box prediction is improved.When the body part of the pedestrian is obscured by other objects,this paper proposes a fusion method of five kinds of Squeeze?and?Excitation blocks and Resden blocks.By re-calibrating the characteristics of the convolution output layer,the visible pedestrian part features are improved in the overall image extraction the proportion of the characteristics.The experimental results show that the multifeature fusion pedestrian detection method for densely occluded pedestrians has a good improvement effect.