Multi-target unmarked attitude estimation method based on deep convolutional neural network

Abstract

The invention discloses a multi-target unmarked attitude estimation method based on a deep convolutional neural network. The multi-target unmarked attitude estimation method comprises the following steps: acquiring an image sequence containing a target behavior from a video; for each image, marking the position and category of each target feature part, and constructing a training data set and a test data set; constructing a multi-target attitude estimation model; training the multi-target attitude estimation model to adjust the weight of the multi-target attitude estimation model, and optimizing the multi-target attitude estimation model; and carrying out precision evaluation on the multi-target attitude estimation model, carrying out target attitude estimation or retraining according to arelation between precision and an expected value, and optimizing the multi-target attitude estimation model. According to the multi-target unmarked attitude estimation method, the accuracy can reachthe human level only through a small number of training images; the target feature part can be effectively tracked under the conditions of messy background, non-uniform illumination, camera distortionand the like; and a good effect can be achieved for different types of targets.

Description

technical field [0001] The invention belongs to the field of computer vision, in particular to a multi-target unmarked pose estimation method based on a deep convolutional neural network. Background technique [0002] The problem of multi-target pose estimation based on computer vision refers to estimating the position and associated information of various parts based on image features. It has a wide range of applications and strong practical value in the fields of human-computer interaction, video surveillance, virtual reality, and automatic driving. However, the research on pose estimation faces many challenges, such as the number of objects in an image cannot be determined in advance, and it is easy for objects to occlude each other in the case of multiple objects. Moreover, as the number of targets increases, the operating efficiency of the algorithm is also facing huge challenges. Before deep learning was widely used, the graph structure (Pictorial Structures) model wa...

AI Technical Summary

Problems solved by technology

However, the research on pose estimation faces many challenges, such as the number of targets in an image is usually not predetermined, and it is easy for targets to occlude each other in the case of multiple targets.

Moreover, as the number of targets increases, the operating efficiency of the algorithm is also facing huge challenges.

Before deep learning was widely used, the graph structure (Pictorial Structures) model was widely used in pose estimation, but the results based on the graph structure model were not very accurate, and it was difficult to extend to multi-target pose estimation

[0003] The proposal of deep learning algorithm provides a new direction for multi-target pose estimation. At present, there are two main ideas for multi-target pose estimation. One is based on the top-down algorithm, which first detects all objects from the image. target, and then use the method of single-target pose estimation to estimate the pose of all targets, but the disadvantages of this algorithm are: the operating efficiency decreases with the increase of the number of targets, and some occluded targets cannot be detected, and the accuracy is not high

However, the current bottom-up pose estimation algorithm is susceptible to interference factors such as uneven illumination, background clutter, and camera distortion in practical applications, and its robustness to interference factors is not high.

And whether it is a bottom-up or top-down approach, the steps of algorithm optimization are relatively complicated, which brings great difficulties to users

Images