Research On Human Detection Based On Deep Learning

With the rapid development of computer vision,visual tasks have permeated so various aspects of our lives,and spawned numerous research topics and applications.As an important branch of object detection,human detection task has attracted more and more attention.It has high application value in intelligent video monitoring,human-computer interaction and autonomous vehicle.In this paper,the human detection method based on deep learning is taken as the research topic,and the following three aspects are analyzed and tested in detail:1.Human detection method based on multi-scale feature fusion.Multi-scale feature fusion combines feature information of different scales.On the one thing,it improves the mismatch between perceptive field and target size caused by detection on a single feature map;on the other thing,it reduces the semantic information differences between feature maps of different scales.Moreover,the shape of anchor was redesigned,and k-means was used to conduct clustering analysis on the annotations of human object dataset,which make it go perfectly for the shape of human object.On the basis of SSD,this paper uses multi-scale fusion to build the overall network structure,and through the redesigned shape of the prior box to match the object information,steadily improve the mean average precision of the model.2.Human detection method based on negative example mining.The one-stage detectors generate anchors of human object through intensive sampling on the feature map,which will result in a large number of simple negative samples,which results in the problem of class imbalance and does harm to the performance of detectors.In this paper,explorering feature information to guide the generation of anchor boxes,which will adaptively generate anchor boxes with different scales and different aspect ratio.Because of the high quality of anchor generation method,it greatly reduces the number of anchor boxes in the background.Then,the Focal Loss function was introduced,which greatly down-weighted the contribution of simple negative samples to the overall loss,with focusing on the hard negatives.It further addresses the classification imbalance problem of the one-stage detectors,improving the performance of the detector in this paper.3.Occluded human detection method based on attention mechanism.The main influence of crowd occlusion is that it significantly increases the difficulty of human localization,which is caused by the shifted predicted bounding box.By imitating the working mechanism of the human visual system,the attention mechanism automatically focuses on the target areas and obtains the key information from the miscellaneous feature information.According to the problem of detecting human in the crowd scene,attention module is embedded in the overall deep learning module,which leads to different weights corresponding to the different location in the feature map.In this paper,Repulsion Loss is used to rebuild the Loss function,adding additional penalty for the predicted bounding box when it shifts to oher surrounding objects.With the proposed method,a more robust human detector is trained end to end.It highly improves the detection accuracy for occlusion cases,showing the effectiveness of this method.