Human body detection and attitude estimation combined deep network learning method

Abstract

The invention discloses a human body detection and attitude estimation combined deep network learning method. The method comprises: combining a CNN model with overall and local information to carry out detection, the model efficiently extracting underlying features through a shared convolution layer, and then the features respectively passing through two branches connected in parallel to carry outhuman body detection and attitude estimation. According to the invention, a Fusion model with a hidden tree structure reasoning algorithm is used to fuse results of human body detection and attitudeestimation, so that a robust and reliable human body detection box is obtained. According to the method, through an NMS algorithm (poseNMS), the obtained information of the human body parts is utilized, and all the individuals which are shielded with one another are effectively reserved. According to the method, a tree structure model is used for embedding information of each part into a detectedbounding box, and a convolutional network is used for realizing an inference algorithm. The method integrates the advantages of overall modeling and local modeling, has a very good detection effect oncrowded and shielded people and pedestrians with uncommon posture behaviors, and can be better integrated into practical application.

Description

【Technical field】 [0001] The invention belongs to the technical field of electronic information, and relates to a deep network learning method combining human body detection and posture estimation. 【Background technique】 [0002] Human detection is an important task in applications such as assisted driving systems and video surveillance. The current human body detector can only achieve better detection results when the human body is unoccluded or not seriously occluded and the human body posture is a common gesture (such as walking). When the human body is detected with gestures and actions (such as jumping over, falling, etc.), the detection effect is not good. This situation inspires the construction of a new human detector to complete the human detection in the above two cases, so as to meet the wider application requirements. [0003] There are two main methods to realize human body detection, one is to model the whole body of the human body as a whole; the other is to...

AI Technical Summary

Problems solved by technology

Although the overall modeling method performs better on general data sets, it is not satisfactory for the detection of crowded and occluded crowds and uncommon postures. The reason is that this overall model is not enough to represent the rich appearance patterns of the human body. ; The local modeling method can better locate the human limbs, but due to the insufficient discrimination between the limbs and the background, this kind of skeleton positioning does not achieve the best effect for human detection, and most of these methods can only achieve a single human detection

Images