CN111814827A - Keypoint target detection method based on YOLO - Google Patents

Abstract

YOLO-based keypoint target detection method, including data set production and processing: On the labeled data set with the original horizontal rectangular frame, add the offset distance (Δx, Δy) of each key point to the upper left corner of the labeled frame , the vertex position coordinates of the upper left corner of the label box are (LUx, LUy), which satisfies that LUx is less than the value in the x direction of all key points, and LUy is less than the value in the y direction of all key points. At this time, the position of each key point is : The fourth quadrant of the coordinate axis when the upper left vertex of the callout box is the origin of the coordinate axis. Point target detection based on the offset of the upper left corner vertex of the prediction frame: The prediction frame is obtained through YOLO, and the offset between each key point and the upper left corner vertex of the prediction frame is obtained at the same time, and the offset corresponding to each key point (Δx) output by the network is obtained. ,Δy) and the coordinates (LUx, LUy) of the upper left corner of the prediction frame are added to obtain the coordinate position of the key point.

Description

Keypoint target detection method based on YOLO

technical field

The invention relates to the technical field of target detection, in particular to a key point target detection method based on YOLO.

Background technique

Visual object detection based on deep learning technology has made great progress in recent years, but there are still many challenging problems. First of all, the current visual target detection model outputs the bounding box of the target, and lacks detection of key points of the target, such as facial feature points in face detection, limb joint points in human body detection, etc. On the other hand, the current target detection algorithm has always been a difficult point for rotating target detection, and the prediction boxes of many target detections are horizontal rectangular bounding boxes. There are two main reasons: 1) Most of the targets in target detection can meet the conditions by using a horizontal rectangular frame, which has a great relationship with the viewing angle of observation. Most of the targets observed from the perspective of people standing are horizontal rectangles. 2) The training of the deep learning model is highly dependent on the annotation of the dataset, and the annotation frame of most datasets is still a horizontal rectangular frame.

P.Objects as Points.arXive-prints，2019：arXiv：1904.07850在文献[1]的基础上添加了中心关键点，用三个关键点来检测目标，提高了准确率和召回率，但其本质仍然是用3个关键点来确定预测框，最终并不输出关键点。With the continuous development of target detection technology, people realize that it is a feasible solution to locate the target through key points, so the literature [1]Law H, Deng J.CornerNet: Detecting Objects as PairedKeypoints[J].International Journal of Computer Vision, 2020, 128(3): 642-656, proposes a method to predict the upper left corner of the target and the lower right corner of the target respectively. The rectangular frame formed by these two key corners is used to locate the target, compared to the center The point prediction method is simpler, but its essence is still to draw a horizontal rectangular box and not output a point target. Reference [2] Zhou X, Wang D,

P.Objects as Points.arXive-prints, 2019: arXiv: 1904.07850 adds the central key point based on the literature [1], uses three key points to detect the target, improves the accuracy and recall rate, but its essence is still It uses 3 key points to determine the prediction frame, and finally does not output the key points.

The Chinese patent (CN201810363952.8) proposes a deep learning-based palm detection and key point location method. The method uses the FasterR-CNN network for training, obtains the palm outline candidate frame and locates the palm key points during detection, and then adjusts the candidate frame threshold. , and select the best palm image with keypoint localization from the candidate frame.

In addition, the detection of rotating and tilting targets also attracts attention. Literature [3] Ma J, Shao W, Ye H, etal. Arbitrary-Oriented Scene Text Detection via Rotation Proposals [J]. IEEE Transactions on Multimedia, 2018, 20 ( 11): 3111-3122. A text detection scheme in any direction is proposed. By setting the anchors-Rotation Anchors with an angle, and then through the RRoI (rotation region of interest) pooling layer, the candidate frame is mapped to the feature map. Go to the classifier to get the result. But RRPN has the problem of being too slow.

Reference [4] Yang X, Liu Q, Yan J, et al. R3Det: Refined Single-Stage Detectorwith Feature Refinement for Rotating Object. arXiv e-prints. 2019: arXiv: 1908.05612. Aiming at the problems of RRPN, RetinaNet is used to construct a single-stage detection framework, and the RefineDet idea is used to refine the first-stage detection results, thereby improving the speed. Chinese patent "201910381699.3" proposes a multi-target detection method for ships based on rotation area extraction. By labeling the rotating target, the final detection result is obtained by calculating the rotation intersection of the preselected frame with the highest confidence and other preselected frames, but The detection accuracy is difficult to guarantee.

Reference [5] Redmon J, Divvala SK, Girshick R, et al. You Only Look Once: Unified, Real-Time Object Detection [C]. computer vision and patternrecognition, 2016: 779-788. (You only look once) yes A single neural network based object detection system proposed by Joseph Redmon and Ali Farhadi et al. in 2015. In order to ensure the efficiency of detection, YOLO proposes the idea of one-stage, which is different from two-stage algorithms such as R-CNN, which need to generate regional proposals, which consumes computing power and leads to slow speed. YOLO does not generate regional proposals, but uses a single convolution. The neural network divides the input image into n*n grids, predicts each grid, directly classifies and regresses the target, and realizes end-to-end detection, so the detection speed is greatly improved. YOLO hits 45fps on the GPU, while its simplified version hits 155fps. Later, in order to improve the accuracy, YOLO proposed YOL09000 and YOLOv3 successively. For example: Literature [6] Redmon J, Farhadi A. YOL09000: Better, Faster, Stronger [C]. IEEE Conference on Computer Vision & Pattern Recognition, 2017: 7263-7271. Reference [7] Redmon J, Farhadi A. YOLOv3: An Incremental Improvement. arXive-prints, 2018: arXiv: 1804.02767.

As a general target detection system with excellent performance, YOLO's speed advantage ensures its feasibility in engineering applications. Therefore, people try to use YOLO to solve related problems, but the original YOLO only outputs horizontal rectangles in target detection. box as the target box. Therefore, the literature [8] Lei J, Gao C, Hu J, et al. OrientationAdaptive YOLOv3 for Object Detection in Remote Sensing Images [C], 2019: 586-597. A method of extending YOLO is proposed to solve the problem of rotating rectangular targets. For the positioning problem, a theta output is added to the output of YOLO, that is, the rotation angle of the predicted frame, but this method can only solve the plane rotation problem of the rectangle. For irregular rectangles, such as trapezoid-like rectangular targets after internal rotation There is still no accurate positioning just by rotating. At the same time, the Chinese patent "CN201910707178.2" also proposes a rotation rectangle target detection method based on YOLOv3. By setting the detection target as a 5-bit vector (x, y, w, h, θ), adding an angle θ, using the belt rotation The anchor point of the angle is used to detect the rotating target. This method can also only deal with the simple rotation of the plane, and still cannot accurately locate the scene such as internal rotation. The Chinese patent "CN201910879419.1" proposes an underwater target detection algorithm based on the improved YOLO algorithm, designs a new loss function, and adds the object aspect ratio information into the loss function, thereby improving the detection of underwater object rotation. The detection effect of rollover and other situations is limited, but the scenes involved are limited. The Chinese patent "CN201910856434.4" proposes a license plate location and recognition method based on the YOLO model. In order to improve the location accuracy of the license plate, an improved YOLO convolutional neural network and a convolutionally enhanced SRCNN (Super Resolution) volume are trained. Integral neural network, in the training of YOLO convolutional neural network, the maxout activation function is used to replace the activation function of the original model, which enhances the fitting ability.

The above improvement method for YOLO improves the ability of the YOLO model to deal with target detection in complex scenes to a certain extent. However, YOLO still has the following problems: 1) In the detection of visual objects with key points, the detection of key points is also important, such as facial features in face detection, limb joint points in human body detection, etc., and YOLO lacks these key points. point detection. 2) There are many irregular rectangles in reality, and rectangular objects with a large aspect ratio under the rotation angle caused by different perspectives, such as license plates of various angles, vehicles photographed in the air, ships and other targets. The prediction frame of YOLO for these rotated and inclined rectangular targets contains a lot of redundant information that is irrelevant to the target.

SUMMARY OF THE INVENTION

In view of the above technical problems, the present invention provides a YOLO-based key point target detection method. By adding a point target detection algorithm on the basis of the original YOLO, YOLO has the ability to detect point targets, so that YOLO can output target detection at the same time. frame and key points, and at the same time achieve precise positioning of rotated rectangular objects in specific applications.

The technical scheme adopted in the present invention is:

The YOLO-based keypoint target detection method includes the following steps:

Step 1. Data set production and processing:

On the annotation dataset whose original annotation frame is a horizontal rectangular frame, add the offset distance (Δx, Δy) of each key point to the vertex of the upper left corner of the annotation frame, and the position coordinates of the vertex at the upper left corner of the annotation frame are (LUx, LUy), which satisfies LUx is less than the value in the x direction of all key points, and LUy is less than the value in the y direction of all key points. At this time, the positions of each key point are: the fourth quadrant of the coordinate axis when the upper left vertex of the label box is the origin of the coordinate axis .

Step 2. Point target detection based on the offset of the upper left corner vertex of the prediction frame:

First, the prediction frame is obtained through YOLO, and the offset between each key point and the upper left corner vertex of the prediction frame is obtained at the same time, and the offset (Δx, Δy) corresponding to each key point output by the network is compared with the coordinates of the upper left corner vertex of the prediction frame ( LUx, LUy) are added to get the coordinate position of the key point.

A YOLO-based key point target detection method of the present invention has the advantages of:

1: While YOLO uses the prediction frame to locate, the distance of each key point from the upper left corner of the prediction frame is predicted in parallel, and then the position of each key point is obtained by combining the position of the upper left corner of the prediction frame. It provides a feasible solution for point target detection.

2: The point target detection scheme of the present invention can deal with the problem of inaccurate positioning and a large amount of redundant information contained in the detection of rotating and inclined rectangular targets. It provides a precise positioning solution for such as inclined license plates, traffic signs, remote sensing detection of high-altitude ships, etc.

3: The point target detection based on the offset of the upper left corner vertex of the prediction frame proposed by the present invention is universal and can be applied to all target detection tasks that require outputting feature points and target frames.

4: The present invention provides a point target detection solution: specifically, the position of the point target is obtained by the offset of the point target from the upper left corner vertex of the initial prediction frame, thereby obtaining the position of the key point. On the basis of the original target detection, by adding point output, the detection capability of the one-stage algorithm is expanded, making it applicable to more scenarios, such as point target detection, accurate positioning of rotating targets, etc.

5: For the specific implementation of the YOLO point target detection scheme, the present invention designs a loss calculation function in YOLO for the offset of the point target relative to the upper left corner vertex of the prediction frame, thereby improving the YOLO loss function. At the same time, the output of YOLO is expanded, so that YOLO outputs the position information of key points while outputting the target frame, which makes the performance of YOLO more powerful and meets more application requirements.

Description of drawings

Figure 1(1) shows that only the frame of the face frame is obtained in the current face detection (lack of detection of key points).

Fig. 1(2) is a schematic diagram showing that the prediction frame contains redundant information in inclined street sign detection.

FIG. 1(3) is a schematic diagram showing that the present invention can simultaneously output a target frame and key points in face detection.

Fig. 1(4) is a schematic diagram of the accurate detection of inclined street signs that can be realized by key point detection in the present invention.

Figure 2 is a schematic diagram of the YOLO target detection process.

Figure 3 shows the relationship between YOLOV3 position prediction and anchor.

Figure 4 shows the unsatisfactory effect of YOLO detecting the rotating rectangular frame.

Fig. 5 (1) is the data set making and algorithm design flow chart of the present invention;

Figure 5(2) is a flow chart of the model detection target of the present invention.

Figure 6 shows the key point position map obtained by offset.

Figure 7 shows the annotation format of YOLOV3 and the key point scheme.

Figure 8 is a flow chart of the specific application of YOLO in rotating rectangle target detection.

Detailed ways

Based on YOLO's key point target detection method, by adding a point target detection algorithm on the basis of the original YOLO, YOLO has the ability to detect point targets, so that YOLO can output target detection frames and key points at the same time, while in specific applications Accurate positioning of rotating rectangular objects is achieved, as shown in Figure 1(1) to Figure 1(4).

Figure 1(1) is a schematic diagram of only obtaining a face bounding box in the current face detection, lacking the detection of key points.

Fig. 1(2) is a schematic diagram showing that the prediction frame contains redundant information in inclined street sign detection, and the effect is not ideal.

FIG. 1(3) is a schematic diagram of simultaneously outputting a target frame and key points in face detection according to the solution of the present invention.

Fig. 1(4) is a schematic diagram of the solution of the present invention realizing accurate detection of inclined street signs through key point detection.

(1): The core idea of YOLO, as shown in Figure 2, YOLO divides the input image into SxS grids. If the actual center point of the object is located in a grid, then this grid will detect the object. The principle of YOLO target detection, here is the principle of YOLOv3 target frame prediction as an example, YOLOv3 first generates 9 anchors through k-means clustering before training, corresponding to the output feature maps of 3 scales, each scale has 3 anchors. Under the network input size of 416, the output feature map sizes of the three scales of YOLOv3 are 13*13, 26*26, and 52*52, respectively, which are used to detect large, medium, and small scale targets respectively. For the feature map at each scale, YOLOv3 gives 3 anchors, and for each pixel grid on the feature map, there will be 3 anchors for prediction, find the most suitable anchor, and give the corresponding offset, which is the prediction frame. YOLOv3 gives 4 values for each prediction box, t _x , _{ty , t w ,} th _h , and the mapping relationship between the 4 values and the final predicted bbox is shown in formulas (1.1) to (1.4).

b _x =δ(t _x )+c _x (1.1)

b _y =δ( _ty )+ _cy (1.2)

Formulas (1.1) to (1.4) are the mapping formulas between the output value and the prediction frame in yolov3.

Among them, t _x and _ty represent the value of the coordinate offset, respectively, _tw and _th represent the scaling of the predicted scale, and _{p w} and ph represent the width and height of the anchor, respectively. δ(t _x ), δ(t _y ) are used to represent the offset of the center point of a certain target relative to the grid responsible for detecting the target, as indicated in Figure 3, where C _x , C _y represent the coordinates of the center point The coordinates of the upper left corner of the grid cell, and the finally obtained b _x , b _x , b _w , and b _h are relative to the center point coordinates and width and height of the feature map. The loss function of YOLOv3 is shown in Equation (1.5).

YOLOv3 loss function includes center coordinate loss Loss _center , formula (1.5.1), width and height loss Loss _wh formula (1.5.2), confidence loss Loss _score formula (1.5.3 ~ 1.5.4), category loss Loss _class formula (1.5.4) Loss of 4 parts in total. The meaning of each variable in formula (1.5) is as follows:

为网格i中第j个边界框的预测。其中各个部分公式中各变量的含义则分别为：公式(1.5.1)λ_coord为动态参数，

为中心坐标的真值，Cxy_i为预测值；公式(1.5.2)中，

和

表示该目标宽度和高度的真实值，w_i和h_i分别表示网络预测该目标的高度和宽度；公式(1.5.3)和公式(1.5.4)，由包含目标时的置信度损失1.5.3和不含目标时的置信度损失1.5.4两个部分构成，其中λ_noobj为不含目标时网络的误差的系数，

和C_i分别代表检测目标的置信度真值和网络预测置信度；式(1.5.5)中

为检测目标概率的真值。Among them, SxS is the number of grids that the network divides the picture, B is the number of bounding boxes predicted by each grid,

is the prediction of the jth bounding box in grid i. The meaning of each variable in each part of the formula is: formula (1.5.1) λ _coord is a dynamic parameter,

is the true value of the center coordinate, and Cxy _i is the predicted value; in formula (1.5.2),

and

Represents the true value of the width and height of the target, w _i and _hi represent the height and width of the target predicted by the network respectively; formula (1.5.3) and formula (1.5.4), the confidence loss when the target is included is 1.5. 3 and the confidence loss 1.5.4 without the target are composed of two parts, where λ _noobj is the coefficient of the error of the network without the target,

and C _i represent the true value of the confidence of the detection target and the confidence of the network prediction, respectively; in formula (1.5.5)

is the true value of the detection target probability.

(2): Point target detection based on the offset of the upper left corner of the prediction frame:

When detecting non-horizontal rectangular targets, the original YOLO obtains the final horizontal prediction frame by predicting the center point and width and height, which leads to the problem shown in Figure 4. In Figure 4, there are two aircraft carriers berthed in the port. The original In an ideal state, YOLO's prediction box is a yellow box, which contains a lot of other information unrelated to the target. At the same time, the prediction boxes of the two aircraft carriers overlap in height. When the non-maximum value suppresses NMS, it is easy to eliminate other predictions and cause missed detection. happened. Such as the technical solution described in the document [9] Neubeck A, Gool L JV. Efficient Non-Maximum Suppression [C]. International Conference on Pattern Recognition, 2006: 850-855. The ideal prediction box should be a red rotated rectangular prediction box, but this is undoubtedly beyond the scope of YOLO's capabilities.

(3): a kind of key point target detection method based on YOLO of the present invention:

Not only is it used to enable YOLO to obtain the ability to detect point targets, but also in specific applications, it can solve the rotating target detection problem shown in Figure 4. After the key points are detected, connect each key point to get a more accurate red prediction frame . The flow chart of the scheme is shown in Figure 5(1) and Figure 5(2): first create a data set, then design a detection algorithm based on the offset of the top left corner of the prediction frame, design a loss function, train the model, and then detect When the point target is obtained by the prediction box and the point offset.

The principle of midpoint target detection in the present invention is shown in FIG. 6 , the number of key points in FIG. 6 is 4, the dotted box in FIG. 6 is the anchor, (p _w , p _h ) is the width and height of the anchor, and the blue The frame is the target prediction frame, the 4 red arrows represent the offset distance of the 4 key points of the target to be detected relative to the top left corner of the prediction frame, and the green frame is the final rotation target frame obtained in the rotation rectangle detection. The formulas for the distance of each key point relative to the offset of the upper left corner of the prediction frame are shown in formulas (1.8) to (2.1).

The model in the present invention outputs t _x , _ty , t _w , _th and 4 sets of offsets for each prediction frame, wherein t _x , _ty , t _w , and _th are used to obtain the prediction frame, that is, It is a blue bounding box, so formulas (1.6)~(1.7) first calculate the width and height of the prediction frame (b _w , b _h ), and then obtain 4 sets of offsets through the prediction frame (b _w , b _h ), As shown in formulas (1.8) to (2.1). Among them, D1 _x , D1 _y are the offset distances of the D1 point relative to the upper left corner of the prediction frame in the x-axis and y-axis directions. Similarly, D2 _x , D2 _y , D3 _x , D3 _y , D4 _x , and D4 _y represent the offset distances from the target key points D2, D3, and D4 to the top-left corner vertex of the prediction frame, respectively.

D1 _X =δ(t _x1 )·b _w D1 _y =δ(t _y1 )·b _h (1.8);

D2 _X =δ(t _x2 )·b _w D2 _y =δ(t _y2 )·b _h (1.9);

D3 _X =δ(t _x3 )·b _w D3 _y =δ(t _y3 )·b _h (2.0);

D4 _X =δ(t _x4 )·b _w D4 _y =δ(t _y4 )·b _h (2.1).

Formulas (1.6) to (2.1) are the calculation formulas based on the algorithm of the vertex offset in the upper left corner of the prediction frame.

The loss function of key points in YOLO point target detection is shown in formula (2.2), and the number of key points in this formula is 4:

If the number of key points increases, the key point loss function will be as shown in formula (2.3), where m is the number of key points:

Formulas (2.2) to (2.3) are the calculation formulas of the offset loss function based on the top left corner vertex offset algorithm of the prediction frame. The present invention adds the calculation loss of key points in the detection of the original YOLOv3, so the final loss of the present invention is The function should be:

Loss _{KeyPoint_offset} = Loss _yolov3 + Loss _KeyPoint (2.4).

The formula (2.4) is the calculation formula of the total loss function based on the vertex offset algorithm of the upper left corner of the prediction frame.

The production and processing of the data set: The key to the processing of the data set in the YOLO point target detection is to add the position information of the key points in the training data set, and add the position information of the key points to the label data set whose original label frame is a horizontal rectangular frame. That is, add the offset distance (Δx, Δy) of each key point to the top left corner vertex of the annotation frame in the annotation data. It should be noted that the vertex position coordinates of the upper left corner of the label box are (LUx, LUy), which must satisfy that LUx is less than the value in the x direction of all key points, and LUy is less than the value in the y direction of all key points. At this time, the positions of each key point are the fourth quadrant of the coordinate axis when the upper left vertex of the callout box is the origin of the coordinate axis. Figure 7 shows the labeling format of the training data in the original YOLO and the labeling format of the keypoint detection scheme, respectively. The label format here is that when the number of key points is 4, when the number of key points increases, the labeling of the corresponding training data set also needs to increase the offset distance of the corresponding key points.

Detection of the model: The point target detection scheme based on the offset of the upper left corner of the prediction frame of the present invention is used to solve the point target detection of YOLO. First, the prediction frame is obtained through YOLO, and the offset of each key point and the upper left corner of the prediction frame is obtained simultaneously. shift. The coordinate position of the key point can be obtained by adding the offset (Δx, Δy) of each key point output by the network to the coordinate (LUx, LUy) of the upper left corner of the prediction frame. In specific applications, such as point target detection, in the detection of facial features, only the position of each key point can be obtained without subsequent processing. When this solution is specifically applied to the detection of rotating rectangular targets, because in the precise positioning of rotating and inclined rectangular targets, it is necessary to draw an accurate positioning frame according to the given four key points. Therefore, the process is shown in Figure 8. The picture is input into the YOLO network, the prediction frame bbox is obtained, and the offset from the key point to the upper left corner of the prediction frame is obtained, which is calculated according to the offset of the upper left corner vertex of the prediction frame and the 4 key points. Find the location of the key points, and then connect the key points to get an accurate red positioning frame.

As a general target detection system with excellent performance, YOLO has good performance in detection accuracy and detection speed, but its horizontal prediction frame cannot provide a better solution for the detection of point targets and oblique rotation rectangular targets Therefore, the present invention enables YOLO to have the ability to detect point targets through the point target detection algorithm based on the offset of the upper left corner of the YOLO prediction frame, and is simultaneously applied in the detection of rotating and inclined targets, thereby achieving the purpose of expanding YOLO.

The present invention proposes a point target detection scheme based on the offset of the upper left corner of the prediction frame. The scheme can output the target frame and feature points at the same time to solve the detection of key points, such as facial feature points, human limb joint points, etc., and at the same time It solves the precise positioning of rotating and inclined rectangular targets such as traffic signs, billboards, and high-altitude remote sensing under oblique viewing angles.

Claims (4)

1. The key point target detection method based on YOLO is characterized in that comprising the following steps: Step 1. Data set production and processing: On the label dataset whose original label frame is a horizontal rectangular frame, add the offset distance (Δx, Δy) of each key point to the upper left corner vertex of the label frame, and the coordinates of the vertex position at the upper left corner of the label frame are (LUx, LUy), LUx is less than the value in the x direction of all key points, and LUy is less than the value in the y direction of all key points. At this time, the positions of each key point are: the fourth quadrant of the coordinate axis when the upper left vertex of the label box is the origin of the coordinate axis; Step 2. Point target detection based on the offset of the upper left corner vertex of the prediction frame: First, the prediction frame is obtained through YOLO, and the offset between each key point and the upper left corner vertex of the prediction frame is obtained at the same time. , LUy) are added to obtain the coordinate position of the key point. 2. the key point target detection method based on YOLO according to claim 1, is characterized in that: In the first step, The number of key points is 4. The formula for the distance of each key point relative to the offset of the upper left corner of the prediction frame is shown in (1.8) to (2.1). The model outputs t _x and _ty for each prediction frame. , t _w , t _h and 4 sets of offsets, t _x , _{ty , t w ,} th _h are used to predict the original target frame, which is the blue bounding frame bbox, so through formulas (1.6) ~ (1.7) first Find the width and height of the prediction frame (b _w , b _h ), and then obtain 4 sets of offsets through the prediction frame (b _w , b _h ), as shown in formulas (1.8) to (2.1);

D1 _X =δ(t _x1 )·b _w D1 _y =δ(t _y1 )·b _h (1.8) D2 _X =δ(t _x2 )·b _w D2 _y =δ(t _y2 )·b _h (1.9) D3 _X =δ(t _x3 )·b _w D3 _y =δ(t _y3 )·b _h (2.0) D4 _X =δ(t _x4 )·b _w D4 _y =δ(t _y4 )·b _h (2.1) Among them: D1 _x , D1 _y is the offset distance of D1 point relative to the upper left corner of the prediction frame in the x-axis and y-axis directions; similarly, D2 _x , D2 _y , D3 _x , D3 _y , D4 _x , D4 _y Respectively represent the offset distances from the target key points D2, D3, and D4 to the top left corner of the prediction frame. 3. the key point target detection method based on YOLO according to claim 2, is characterized in that: the loss function of key point in YOLO point target detection is as shown in formula (2.2), and the number of key points in this formula is 4:

If the number of key points increases, the key point loss function will be as shown in (2.3), where m is the number of key points:

YOLO point target detection is to increase the calculation loss of key points in the original YOLOv3 detection, so the final loss function is: Loss _{KeyPoint_offset} = Loss _yolov3 + Loss _KeyPoint (2.4). 4. the key point target detection method based on YOLO according to claim 2, is characterized in that: In the second step, the picture is input into the YOLO network, the offset from the key point to the upper left corner of the prediction frame is obtained while the prediction frame is obtained, and the key point is calculated according to the offset of the upper left corner vertex of the prediction frame and the four key points. The position of the point, and then the key points are connected to obtain an accurate positioning frame.

Images