Motion Prediction Using Depth Information of Human Arm Based on Alexnet

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019
DOI: 10.5121/ijaia.2019.10402 9
MOTION PREDICTION USING DEPTH
INFORMATION OF HUMAN ARM BASED ON
ALEXNET
JingYuan Zhu1
, ShuoJin Li1
, RuoNan Ma2
, Jing Cheng1
1School of Aerospace of Engineering, Tsinghua University, Beijing, China
2
School of Economics and Management, Tsinghua University, Beijing, China
ABSTRACT
The development of convolutional neural networks(CNN) has provided a new tool to make classification
and prediction of human's body motion. This project tends to predict the drop point of a ball thrown out by
experimenters by classifying the motion of their body in the process of throwing. Kinect sensor v2 is used to
record depth maps and the drop points are recorded by a square infrared induction module. Firstly,
convolutional neural networks are made use of to put the data obtained from depth maps in and get the
prediction of drop point according to experimenters' motion. Secondly, huge amount of data is used to train
the networks of different structure, and a network structure that could provide high enough accuracy for
drop point prediction is established. The network model and parameters are modified to improve the
accuracy of the prediction algorithm. Finally, the experimental data is divided into a training group and a
test group. The prediction results of test group reflect that the prediction algorithm effectively improves the
accuracy of human motion perception.
KEYWORDS
Human Motion, Prediction, Convolutional Neural Network, Depth Information
1. INTRODUCTION
In recent years, more and more researches on human motion classification have been carried out.
Some of them are mainly based on 3D skeleton while others are based on depth sequences.
Identification, classification and prediction of human motion play significant roles in monitoring,
human-machine interaction, body language application. The development of computer vision and
image processing technology provides great room for its improvement as well. Besides, some
artificial intelligence methods have also been put into use for this field.
Development and success during the past decade have already proved the effectiveness and
availability in the field of computer vision. Sherif Rashad et al. [1] implemented a mobile
application with machine learning method to improve security of mobile devices with predicting
users’ behaviour. Dilana et al. [2] applied machine learning methods including k-Nearest-
Neighbors and Linear Discriminant Analysis to emotion classification. Abdelkarim Mars [3]
proposed an Arabic online recognition system with neural network composed of Time Delay
Neural Networks (TDNN) and multi-player perceptron (MLP). Alayna Kennedy et al. [4] made
use of artificial neural networks to analyze information contained in electromyogram signals to
classify human walking speed. In our work convolutional neural network is applied to make
classification of human motion. This paper proposed a human motion classification model based
on AlexNet with considerable correctness.

10
2. RELATED WORK
Johansson’s [5] motion-sensing experiment which was conducted in 1970s was considered as one
of the pioneers in this field. In their experiment, several bright spots distributed on the human
body are used to study human motion perception. Their research proves that human vision can not
merely detect motion direction, but also detect different types of limb movement patterns,
including recognition activities and different motion patterns. Many researches dealt with topics
such as surface treatment, curves and contours, pattern recognition and feature extraction [6].
Arridhana Ciptadi et al. [7] presented a novel action representation based on encoding the global
temporal movement of a human action. They found that temporal dynamics of an action are
robust making it useful for action recognition and retrieval. Mosabbeb E et al. [8] proposed a
distributed activity classification framework, in which they make use of several camera sensors to
observing the scene. Their method could also be applied to human activity classification. Nicola
Bellotto et al. [9] applied multi-sensor data fusion techniques and used the onboard laser range
finder (LRF) for laser-based leg detection. They performed human tracking under complex
environments.
Due to the progress of depth cameras, depth images are more and more used in motion prediction
and recognition. Shotton J et al. [10] estimated the joint position of human based on the depth
images. According to this method, the posture of human body could also be defined by the
position of joint position. Ellis et al. [11] used a latency-aware learning formulation to train a
logistic regression-based classifier that automatically determines distinctive canonical poses from
data and uses these to robustly recognize actions in the presence of ambiguous poses such as
balance and kick poses. Lu Xia [12] proposed a model-based approach detecting humans using a
2-D head contour model and a 3-D head surface model with data collected by Kinect sensor.
Apart from using the data of skeleton, depth maps have also been applied to this field for a period
of time. They are widely used especially in scene or object reconstruction and robots. Depth maps
record the distance from the camera at a corresponding point in the scene in each pixel. Foix, S. et
al. [13] introduced a 3D time-of-flight (ToF) camera which illuminates the scene with a
modulated light source and observes the reflected light. The Kinect sensor (version 2) belongs to
this kind of camera.
Many approaches applied to the field of gesture recognition are based on computer vision which
is different from approaches detecting 3D skeleton coordinates or recognizing the whole body
action stated above. In 1978, Stokoe [14] used four aspects including hand shape, position,
orientation and movement to represent gesture. Many application systems related to gesture
recognition have been put into use. Imagawa, K. et al. [15] proposed a system which could
automatically analyze and annotate video sequences of technical talks. The system would track
and recognize gestures to provide annotation. Triesch, J, et al. [16] presented a person-
independent gesture interface implemented on a real robot allowing users to give simple
commands to realize human-robot interaction. Nolker et al. [17] detected 2D location of fingertips
by Local Linear Mapping neural network and mapped them to 3D position with which method
they could recognize hand pose under different views.
A lot of related works on full-body movements have also been carried out. Bobick [14] presented
several approaches to the machine perception of motion and discusses the role and levels of
knowledge in each. Yang Wang et al. [18] considered the problem of describing the action being

11
performed by human figures of still images. Ji, Shuiwang et al. [19] developed a novel 3D CNN
model for action recognition.
AlexNet is a convolutional neural network model designed by ImageNet competition winner
Hinton and his student Alex Krizhevsky [20]. Due to the superiority of AlexNet, it has been
applied to many fields of classification. For example, Yuan, Z. W. et al. [22] had applied
AlexNetto improve the result of image retrieval with its fusion feature. Shenshen Gu et al. [23]
applied this model to test whether there is a tennis ball in the picture. Jing Sun et al. [24]
improved the accuracy of the scene image classification with their model based on AlexNet. The
efficiency of AlexNet has also been tested with some datasets. Pal, Raju et al. [25] applied
AlexNet to Animal Diagnostics Lab (ADL) dataset to extract features. Y Dang et al. [26] did their
study on the evaluation of land cover classification using remote sensing images based on
AlexNet.
It should be noted that most of the classification researches are based on large amount of accurate
calculation based on kinematics data which requires massive amount of calculation. Besides, it
would be difficult to make predictions of human motion. Taking these disadvantages into
consideration, we tend to make use of convolutional neural networks(CNN) to analyze human
kinematics data obtained from depth maps shot by Kinect sensor(version 2) and make prediction
with an appropriate algorithm. The prediction model is mainly based on AlexNet trained with
depth information.
3. PROJECT AND ALGORITHM
3.1. Project Structure
Ball throwing is chosen as the process to make analysis and prediction of in this project. The
objective of the algorithm is to make analysis of the process of throwing ball and predict its drop
point. Ball is thrown into the area which is divided into 4 parts of the same area size.
Figure 1. Project Flowchart
The human motion data is recorded in 30 frames depth maps with Kinect sensor v2. To increase
effectiveness and shorten the time cost in training and predicting, those data would be
preprocessed to keep the most important part and ignore other parts which have few impacts on
training and prediction. Then they would be divided into two groups including training data and
testing data for network model modified from AlexNet. The structure and parameters of the
network would be adjusted to improve accuracy of drop point prediction. With these methods we
could downsize the input data size and increase the efficiency and accuracy of the algorithm.

12
3.2. Prediction Algorithm
Convolutional neural networks model is made use of to analyze data obtained from depth maps.
The algorithm is modified from AlexNet which had been proved available in many fields.
Compared with other classical machine learning classification algorithms, the AlexNet has many
advantages which would help it perform better in classification. Firstly, AlexNet uses RELU as
the activation function of CNN whose effect would be better for deeper networks. Secondly, to
avoid the problem of over-fitting while training the model, AlexNet takes advantage of dropout
layers to randomly ignore part of changes in neurons. Besides that, the usage of overlapping max-
pooling layers which are different from the average-pooling layers that had been widely used in
CNN before could avoid the fuzzification effect and improves the richness of features at the same
time. In AlexNet model, the addition of LRN layer could create a competition mechanism for the
activity of local neurons under which would make relatively larger values become larger and
correspondingly the neurons with lower feedback would be suppressed. All of these
improvements could help AlexNet model to make more accurate classification.
The original structure of AlexNet includes five convolution layers, two overlapping max-pooling
layers distinctively following the first two convolution layers and two fully connected layers. The
output of the last fully connected layer will then be put into the Softmax classifier to get the final
result.
In the process of analysis, we tried to modify the network model through methods including
adding dropout layers into our model and adjust loss function to obtain better classification results.
The dropout layer could reduce the severity of over-fitting while different loss functions could
lead to different results so an appropriate loss function would help us train the model with
reasonable loss values. Apart from those, it would be significant to find a matched scale of
network. Parameters such as learning rate and batch size also need adjustment.
4. EXPERIMENT
4.1. Laboratory Equipment
Experimenters are students picked out from our lab. The whole experiment was also carried out in
our lab as well. To make sure that the effect of the model wouldn’t be seriously limited by the
amount of our data, we repeated the experiments 400 times to get depth image data needed for
network training.
In the experiment, the equipment includes a Kinect sensor v2 with a tripod for fixation, a square
infrared module whose area is 2.25 square meters, several computers to control the sensor and
module. Apart from those, a ball is needed for experimenters to throw.
Kinect sensor v2 is used to record the depth maps in the process of experimenters’ throwing balls.
It is controlled by a C++ program to record the first 30 frames of depth maps since the
experimenter’s right arm is raised higher than his waist. The square infrared module is controlled
by algorithm coded by C++ to record the first drop point of falling ball in every experiment. The
module could detect not only the drop area but also the exact coordinate of the drop point
according to which the area of drop point could be calculated.

13
In the experiment of data collection, Kinect sensor v2 is placed to record the process of
experimenters’ body motion, especially their right arms. Distance between experimenters and the
square infrared module’s center is 195 centimeters. Experimenters are asked to stand with their
feet separated at a distance of 30 centimeters. Kinect sensor v2 is fixed at the height of 135
centimeters. Distance between Kinect sensor v2 and the module is 27 centimeters. Distance
between experimenters and the module is 95 centimeters. With different parameters, the networks
model would provide us with various results.
4.2. Experimental Methods
The experimenters are asked to throw the ball with their right arms into an area which is divided
into four parts of the same size with similar postures. The main goal of the algorithm is to analyze
the process of ball throwing and predict the drop point of the ball according to motion of human
body, mainly human arms.
All of the data collected by Kinect sensor are recorded in depth maps. Kinect sensor is activated
whenever the assistant’s hand is higher than his waist and would begin to record the 30 frames.
The size of every frame is 217088 while the length of one frame is 512 and the width is 424. The
value of every point is normalized to the range from 0 to 255. Taking the great size of every
frame and that even more than 50 per cent of a frame is actually empty without any information
and that the drop point is mainly determined by the motion of experimenters’ upper bodies into
consideration, those frames are cropped to a relatively smaller size with the width of 110 and
height of 129. After pre-processing, experimenters’ bodies would take up more than 70 per cent
space of a picture.
In the process of throwing, experimenters first lift their right arms from the front of their bodies
and throw the ball out when the right hands are at almost the same height as their overhead.
Experimenters are asked to try their best to keep other parts of their body steady enough to
guarantee that the fall of the ball is only influenced by the motion of their arms.
Figure 2. Photo of Experiments

14
Figure 3. Kinect Sensor and Infrared Module
5. EXPERIMENTAL RESULTS AND DISCUSSION
The 30 frames of depth images recorded in one experiment also contain experimenters’ body
motion information after the ball is thrown away. With observation and algorithm testing, the first
10 frames recorded since the experimenters’ right hands are raised higher than their waists were
finally decided to be the data put into the network model.
Figure 4. Photo and Depth Image of Throwing Ball
The convolutional neural networks model based on AlexNet is firstly trained and tested with data
collected from our experiments. Result of prediction accuracy obtained through cross-validation
is shown in Figure 5. The accuracy of prediction is slightly over 50 percent which can’t be put
into use in reality. Therefore, modifications of parameters and structure of networks are needed to
increase effectiveness and availability of the action classification algorithm.

15
Figure 5. Prediction Result 1
As for the convolutional neural networks, it is important to decide the scale of the net which is
closely linked to effectiveness and accuracy of training and predicting. Since the depth images
have much in common, the neural networks model has great possibility to meet the problem of
over-fitting which would badly effect the model’s output.
To avoid the problem of over-fitting, we downsize the network model to a relatively proper one.
After enough adjustment and experimenting, we finally confirmed the number of neurons in the
net which is appropriate for input data.
Loss function plays an important role in CNN training. Neural networks model’s training process
is to minimize the value of loss function. At first the loss function only contains MSE between
truth and prediction. To further avoid the model from effect of over-fitting, modification of the
loss function by adding regular terms was carried out. The regular term chosen is L2
regular term
which is shown as below. This modification proves useful for alleviating over-fitting and
improving the model’s accuracy.
With regular terms added into the loss function, the results of model’s prediction accuracy were
increased further as shown in Figure 6.

16
The normalization layers are often added to networks model in order to control the cost of the
training process. However, under some circumstances, normalization layers would actually
decrease the accuracy of neural network models. Normalization layers sometimes effects the
calculation of weight in training processes since it will decrease some parameters’ influence on
the network output. Frames put into the model has the width of 110 and height of 129, value of
every point varies from 0 to 255 reflecting depth information. Values of loss function are finite in
training. After analyzing and testing, normalization layers are aborted to obtain higher accuracy.
Adding dropout layers is a common way of preventing over-fitting. Dropout layer’s existence
makes some of the implicit layer nodes in the networks stop working randomly during the
training process. Dropout layers could reduce the interaction between feature detectors which
could improve the effectiveness and accuracy of the model.
Dropout layers are added into the model with various sequences to test the accuracy of training
and accuracy. After lots of attempts, it is found that adding the dropout layers after the first max-
pooling layer and the second max-pooling layer could give out the best result which is shown in
Figure 7.

17
After modification made to the convolutional neural network model, the most appropriate
structure of convolutional neural networks model to make prediction of drop points is acquired.
The networks structure is shown in Figure 8.
Figure 8. Model Structure
Other parameters such as learning rate and batch size are also adjusted. With this modified model,
the accuracy of prediction would be higher than 80 per cent. The high level of accuracy for
prediction proves that the model is able to be applied to the field of predicting human’s arm
motion.
6. CONCLUSION AND FUTURE WORK
This paper proposed a model to make classification of human motion based on AlexNet. To make
the model designed by us reliable enough, the convolutional neural network model is trained with
depth images data of human motion obtained from about 400 times of experiments. We made
several modifications to the convolutional neural network model to make it appropriate for the
target we want to reach. The classification results showed that our convolutional neural network
model could be applied to make classification of human motion and make prediction of ball
throwing. With accuracy our model has reached, it could be used to catch things like balls thrown
out by people practically. Apart from that, our model could also be applied to other fields
including human-machine interaction, video monitoring, body action language and medical care.

18
In future work we would try to make prediction from different angles by adding more images
recorded from different angles. Besides, limited by size of data, drop area is only divided into
four parts so far. It is possible for the network model to make prediction of drop area which is
divided by finer meshes with larger amount of data. All depth images recorded since
experimenters raised their right hands higher than their waists are used as training data without
distinguishing. They would take up the same weight in training and testing. Surely it may be more
reliable and reasonable to distribute those frames of suitable weights before training the model. In
conclusion, enrichment of dataset and improvement of algorithm will help us gain better results.
ACKNOWLEDGEMENTS
In the whole process of this project, we got adequate support from Tsinghua University. We want
to express our thanks for Dr. Ou Ma for his help and guidance.
REFERENCES
[1] Sherif, R. and Jonathan, B., 2018. “Behavior-based security for mobile devices using machine
learning techniques”. International Journal of Artificial and Applications(IJAIA), 9(4).
[2] Dilanam, H., Lin, Z., and Harald C. 2018. “Performance evaluation of various emotion classification
approaches from physiological signals”. International Journal of Artificial and Applications(IJAIA),
9(4).
[3] Abdelkarim Mars and Georges Antoniadis. 2016. “Arabic online handwriting recognition using neural
network”. International Journal of Artificial and Applications(IJAIA), 7(5).
[4] Alayna Kennedy and Rory Lewis. 2016. “Optimization of neural network architecture for
biomechanicclassdfication tasks with electromyogram inputs”. International Journal of Artificial and
Applications(IJAIA), 7(5).
[5] Johansson, G., 1973. “Visual perception of biological motion and a model for its analysis”.
Perception Psychophysics, 14(2), pp. 201–211.  
[6] IEEE, 2002. “Proceedings of ieee conference on computer vision and pattern recognition”. IEEE
Computer Society Conference on Computer Vision Pattern Recognition, Cvpr.  
[7] CiptadiA.,Goodwin, M.S.,andRehg,J.M.,2014.“Movement pattern histogram for action recognition
and retrieval”. European Conference on Computer Vision.  
[8] Adeli, M. E., Raahemifar, K., and Fathy, M., 2013. “Multi-view human activity recognition in
distributed camera sensor networks”. Sensors, 13(7), pp. 8750–8770.
 
[9] Nicola, B. and Huosheng, H., 2009. “Multisensor-based human detection and tracking for mobile
service robots”. IEEE Transactions on Systems Man Cybernetics Part B Cybernetics A Publication of
the IEEE Systems Man Cybernetics Society, 39(1), pp. 167–81.  
[10] Shotton, J., Fitzgibbon, A., Cook, M., Sharp, T., Finocchio, M., Moore, R., Kipman, A., and Blake,
A., 2013. Real-Time Human Pose Recognition in Parts from Single Depth Images.  
[11] Ellis, Chris, Masood, Zain, S., Tappen, Marshall, F., LaViola, Joseph, J., Jr, and Sukthankar, 2013.
“Exploring the trade-off between accuracy and observational latency in;action recognition”.
International Journal of Computer Vision, 101(3), pp. 420–436.  
[12] Lu, X. L. X., 2011. “Human detection using depth information by kinect”. IEEE Computer Society
Conference on Computer Vision Pattern Recognition, Cvpr.

19
[13] Foix, S., Alenya, G., and Torras, C., 2011. “Lock-in time-of-flight (tof) cameras: A survey”. IEEE
Sensors Journal,   11(9), pp. 1917–1926.  
[14] Stokoe, W. C., 1980. “Sign language structure”. Annual Review of Anthropology, 9(9), pp. 365–390.
[15] Imagawa, K., Lu, S., and Igi, S., 1998. “Color-based hands tracking system for sign language
recognition” IEEE International Conference on Automatic Face & Gesture Recognition.  
[16] Triesch, J., and Malsburg, C. V. D., 1999. “A gesture interface for human-robot-interaction”.   IEEE
International Conference on Automatic Face & Gesture Recognition
[17] Nlker, C., Ritter, H., Fakultt, T., and Bielefeld, U., 1998. “Illumination independent recognition of
deictic arm pos-   tures”. In Conference of the IEEE Industrial Electronics Society.
[18] Bobick,A.F.1997.“Movement, activity and action:the role of knowledge in the perception of
motion”.Philosophical Transactions of the Royal Society of London, 352(1358), pp. 1257–1265.
[19] Wang, Y., Jiang, H., Drew, M. S., Li, Z. N., and Mori, G., 2006. “Unsupervised discovery of action
classes”. IEEE Computer Society Conference on Computer Vision Pattern Recognition.
 
[20] Shuiwang, J., Ming, Y., and Kai, Y., 2013. “3d convolutional neural networks for human action
recognition”. IEEE Transactions on Pattern Analysis Machine Intelligence, 35(1), pp. 221–231.
[21] Krizhevsky A, Sutskever I , Hinton G. 2012. “ImageNet Classification with Deep Convolutional
NeuralNetworks”. Advances in neural information processing systems, 25(2).
[22] Yuan, Z. W., and Zhang, J., 2016. “Feature extraction and image retrieval based on
alexnet”.EighthInternational Conference on Digital Image Processing.
[23] Gu, S., Lu, D., Yue, Y., and Chen, X., 2017. “A new deep learning method based on alexnet model
and ssd model for tennis ball recognition”. IEEE International Workshop on Computational
Intelligence Applications.  
[24] Jing, S., Cai, X., Sun, F., and Zhang, J., 2016. “Scene image classification method based on alex-net
model”. International Conference on Informative Cybernetics for Computational Social Systems. 
[25] Pal, R., and Saraswat, M., 2018. [IEEE 2018 Eleventh International Conference on Contemporary
Computing (IC3) - Noida, India (2018.8.2-2018.8.4)] 2018 Eleventh International Conference on
Contemporary Computing (IC3) - Enhanced Bag of Features Using Alexnet and Improved
Biogeography-Based Optimization for Histopathological Image Analysis. pp. 1–6.  
[26] Dang, Y., Zhang, J., Deng, K., Zhao, Y., and Fan, Y. U., 2017. “Study on the evaluation of land cover
classification using remote sensing images based on alexnet”. Journal of Geo-Information Science,
19(11).

Motion Prediction Using Depth Information of Human Arm Based on Alexnet

More Related Content

What's hot (17)

Similar to Motion Prediction Using Depth Information of Human Arm Based on Alexnet (20)

More from gerogepatton (20)

Recently uploaded (20)

Motion Prediction Using Depth Information of Human Arm Based on Alexnet