IEEE Robotics & Automation Magazine - December 2023 - 70

fast speed and high precision [18], which was selected as the
basic model for cow face detection.
The network architecture of YOLOv5s consists of backbone,
neck, and head. The backbone part is composed of
cross-stage partial network (CSPNet) and spatial pyramid
pooling (SPPF), which map the features of the base layer into
two parts, and carry out feature fusion through local crosslayer
structure. The neck part uses a path aggregation network
(PANet) to effectively aggregate the semantic information in
the high level, and then fuses the shallow-layer features from
the bottom up. The head is used to make a final prediction
based on the features obtained from the neck layer.
In this study, the cow face detection and recognition results
are output. Although YOLOv5s is feasible to extract cow face
Backbone
Conv
Conv
C3
CA
Conv
C3
CA
Conv
C3
CA
Upsample
Conv
C3
CA
SPPF
Detection Box
Position
Embedding
Class Token
Transformer
Encoder
DeepSORT-ViT MULTIOBJECT
TRACKING
Based on YOLOv5s-CA detection
and recognition of cow face, the
multiface tracking method was further
studied. DeepSORT is an online
target tracking algorithm, which
mainly includes track state estimation,
association, and cascade matching
[12]. The information on track
state estimation is defined in the 8D
state space (,xc
h ).o
y ,c
c , h,
x ,
.
c
y ,
.
c
c ,o
The first four variables are the
DeepSORT
Track
ReID
Output
FIGURE 1. YOLOv5s-CA+DeepSORT-ViT base on cow tracking and feeding monitoring.
70 IEEE ROBOTICS & AUTOMATION MAGAZINE DECEMBER 2023
coordinates of the center of the face
frame, aspect, ratio, and height, respectively;
the last four variables are the
change rate of the corresponding
variables. The KF algorithm is used
to predict the position and state of
the target in the next frame. For
association and cascade matching,
the Hungarian algorithm is used to
Conv
C3
YOLO
HEAD
Concat
C3
Concat
Upsample
Conv
Concat
C3
Conv
YOLO
HEAD
Neck
C3
Concat
Conv
where
gic
h
() and
gjc
w
() represent the
attention weight matrix of H and W,
respectively. Finally, the acquired
image (, )yi jc
enhances the sensitivity
to position information. As shown in
Figure 2, CA was added to each C3
module in the backbone part of
YOLOv5s, and (, )yi jc
passed through
the neck and head parts to output the
cow face detection results for subsequent
tracking.
YOLO
HEAD
features, the complex background (e.g., pens, crush, forages)
brings visual occlusions, which affect the face detection performance.
To further focus on the cow face-related features,
CA was added [19] in the YOLOv5s.
As shown in Figure 2, CA embeds coordinate information
into channel attention, then generates the so-called CA. For
input face feature X, the pooled kernels of dimensions (H, 1)
and (1, W) are used to encode each channel along horizontal
and vertical coordinate directions. The two spatial direction
coding output expressions are as follows:
zh 11 xj w(, ).
w
W 1
-
c
h
()
==
=
W 0
i
xh iz w
c
c(, ), ()
H 1
-
// (1)
Then the output tensors of the two
H i =0
c
spatial directions () and ()
zhc
h
Prediction
zwc
w
are splicing, using 11# convolution.
The final acquired facial feature output
is:
yi jx ij gi gjcc c
h
(, )(,)
= ## w
()
c () (2)
Patch Embedding
IEEE Robotics & Automation Magazine - December 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2023
