Study On Human And Head Joint Detection Based On Deep Learning

Object detection is one of the research hotspots in the field of computer vision.In object detection,human detection and head detection based on deep learning have shown a wide range of application prospects and needs.For example,detecting and identifying a person requires not only detecting the human body,but also detecting the head to obtain the information of the head for further recognition and judgment.There is a certain spatial position relationship between the human body and the head.The joint detection of the human and the head through spatial location modeling can effectively improve the accuracy of the detection of the human and the head.This paper takes the extracted features of detected object and the generated proposals of detected object in object detection as the starting point.Taking deep learning as the method and focusing on the joint detection of human and head.Using the similar and unique features of human and head,and at the same time combining the inclusion relationship of human and head.The main research contents and innovations of this thesis are as follows:1.We propose a feature enhancement human and head joint detection.This method performs different processing on the extracted features of human and head,and considers the similar and unique features of the human and head and designs two different module branches.One module is the human enhancement module,which enhances the extracted features,and the enhanced features are more conducive to detecting the human.The other module is the head enhancement module,which also enhances the extracted features,and the enhanced features are more conducive to detecting the head.This is more effective for detecting human and head.2.We propose a joint detection of human and head relationship regression.This method optimizes the generated human and head proposals and contains two modules.The first module is the method of human and head proposals relationship generation.Considering that the human contains the head,the human proposals are first extracted through the region proposal network,and then the proposals of the head are found in the human proposals by learning the relationship between the human and the head.The second module is the human and head proposals asymptotic regression method.The human and head proposals extracted by the first module are performed border regression twice,and continuously improve the quality of proposals,so as to improve the performance of human and head joint detection.3.We propose a feature enhancement multi-scale joint detection of human and head relationship regression.The method simultaneously improves the extracted human and head features and the generated human and head proposals.Considering the similar and unique features of the human and the head,and the inclusion relationship between the human and the head,combining the feature enhancement method and the proposals relationship regression method.The feature enhancement method is used to enhance the features of the human and the head separately.At the same time,the human and head proposals extracted by the RPN and a multi-scale proposals module is added to scale and translate the proposals of the human and the head,and then use the proposals relationship regression method to perform twice border regressions to achieve better performance of joint detection of the human and the head.This paper proposes three types of joint detection of human and head based on deep learning,explores and verifies the effectiveness of each method,and proves the feasibility of joint detection of human and head based on deep learning,and improves performance for detecting human and detecting head.