Can bring the ideal effect.

Human posture estimation is an important research direction in the field of computer vision, which is widely used in human activity analysis, human-computer interaction and video surveillance. Human posture estimation is to specify the key points of human body (such as shoulder, elbow, wrist, hip, knee, ankle, etc.). ) in an image or video by a computer algorithm. This paper mainly introduces the development of human posture estimation methods after the rise of deep learning in recent years.

Second, what is the use of human posture estimation?

(1) Use human posture to detect falls or enhance safety and monitoring;

(2) used for fitness, sports and dance teaching;

(3) Training the robot to "learn" to move joints;

(4) Tracking the movement of human body in movie special effects production or interactive games. By tracking the change of human posture, the fusion and synchronization of virtual characters and real characters are realized.

Thirdly, the evaluation index of human posture estimation algorithm.

(1)OKS (similarity of key points of objects)

OKS is an evaluation index put forward by COCO Attitude Estimation Challenge. The COCO ranking shows that the highest map of Challenge 18 is 0.764. Maps based on similarity of key points of objects;

Where di represents the Euclidean distance between the predicted key point and ground truth; Vi is the visibility symbol of the truth on the ground; S is the target scale, which is equal to the square root of human area in the ground truth value; Ki controls every key constant of attenuation.

(2)PCK (probability of correct key point)

The evaluation index of MPII data set is PCKh@0.5, and the highest PCKh of MPII data set is 92.5 at present. When the normalized distance between the predicted joint point and its corresponding real joint point is less than the set threshold, it is considered that the joint point is predicted correctly, and PCK is the proportion of the joint points predicted correctly by this method.

PCK@0.2 refers to the diameter of the trunk as a reference. If the normalized distance is greater than the threshold value of 0.2, the prediction is considered to be correct.

PCKh@0.5 means that the head length is taken as a reference, and if the normalized distance is greater than the threshold value of 0.5, the prediction is considered correct.

(3)PCP (percentage of correct parts)

If the critical distance between the position of two joint points and the real limb is at most half of the length of the real limb, it is considered that the joint points are correctly predicted, and PCP is the proportion of the joint points correctly predicted by this method.

Fourthly, the development of human posture estimation algorithm.

In 20 13, Toshev and others introduced DeepPose into the field of human posture estimation, and the research of human posture estimation began to shift from traditional methods to deep learning. The following will summarize six works that I personally think are iconic in chronological order.

(1)DeepPose(20 14, Google)

DeepPose, proposed by Alexander Toshev and Christian Szegedy, applies CNN (Convolutional Neural Network) to joint detection of human body for the first time. DeepPose transforms human posture estimation into joint point regression, and puts forward a method of applying CNN to human joint point regression: input the whole image into 7-layer CNN for joint point regression, and further use cascaded CNN detectors to improve the accuracy of joint point positioning.