计算机代考 CSCI473/573 Human-Centered Robotics

CSCI473/573 Human-Centered Robotics
Project 3: Robot Understanding of Human Behaviors Using Skeleton-Based Representations
Project Assigned: April 13
Multiple due dates (all materials must be submitted to Canvas) Deliverable 1 (Representation Code) due: April 22, 23:59:59 Deliverable 2 (Complete Code) due: April 29, 23:59:59 Deliverable 3 (Project Report) due: May 04, 23:59:59

Copyright By PowCoder代写 加微信 powcoder

In this project, students will implement several skeleton- based representations (Deliverable 1) and use Support Vector Machines (SVMs) (Deliverable 2) to classify human behav- iors using a public activity dataset collected from a Kinect V1 sensor. Additionally, students are required to write a report following the format of standard IEEE robotics conferences using LATEXin Deliverable 3.
Students are required to program this project using C++ or Python in Ubuntu 18.04 LTS, but students are NOT required to implement the project in ROS.
Before you start this project, you need to understand the related content in lecture slides of Chapter 09: “09-Skeleton- based Representations” (for Deliverable 1) and Chapter 10 “10-Robot Learning from Data” (for Deliverable 2), as well as the corresponding lecture videos on the course website.
I. DATASET
The MSR Daily Activity 3D dataset1 will be used in this projet, which was one of the most widely applied benchmark dataset in human behavior understanding tasks. This dataset contains 16 human activities, as demonstrated in Figure 2, performed by 10 human subjects. Each subject performs each activity twice, once in a standing position, and the other in a sitting position. Although this dataset contains both color- depth and skeleton data, in this project, we will only explore the skeletal information to construct skeleton-based human representations (Deliverable 1).
The skeleton in each frame contains 20 joints, as illustrated by Figure 1. The correspondence between joint names and joint indices is also presented in the figure. For example, Joint #1 is HipCenter and Joint #18 is KeeRight, etc.
In this project, a pre-formatted skeleton data will be used, which can be downloaded in the course website. http://inside.mines.edu/ ̃hzhang/Courses/ CSCI473-573/assignment.html.
This dataset in this project is a subset of original dataset, which only contains six (6) activity categories:
This write-up is prepared using LATEX.
1The MSR Daily Activity 3D dataset is removed from the author’s website and no longer publicly available after 2017.
Fig. 1: Skeleton joint names and indices from Kinect SDK.
• CheerUp (a08)
• TossPaper (a10)
• LieOnSofa (a12)
• Walk (a13)
• StandUp (a15)
• SitDown (a16)
It is also noteworthy that this dataset is not formatted to the LIBSVM format but much easier to process then the original data format. Conversion of the data to the SVM format will be part of Deliverable 2.
In particular, the dataset contains two folders: Train and Test. Data instances in the directory Train is used for training (and validation). Data instances in Test is used for testing the performance. Each instance has a filename like: a12 s08 e02 skeleton proj.txt. This filename means the data instance belongs to activity category 12 (i.e., a12, that is “lie down on sofa” as in Figure 2), from human subject 8 (s08) at his/her second trial (i.e., e02). The dataset contains 16 activity categories, 10 subjects, and 2 trials each subject. Instances from subjects 1–6 are used for training (in the directory Train), and instances from subjects 7–10 are used for testing (Test).
When you open a data instance file, e.g., a12 s08 e02 –
(b) Indices

skeleton proj.txt, you will see the following:
Algorithm 1: RAD representation using star skele- tons
Input : Training set Train or testing set Test Output : rad d1 or rad d1.t
for each instance in Train or Test do forframet=1,…,T do
Select joints that form a star skeleton (Figure 3); Compute and store distances between body
extremities to body center (dt1,…,dt5); Compute and store angles between two adjacent
body extremities (θ1t , …, θ5t );
Compute a histogram of N bins for each di = {dti}Tt=1, i = 1,…,5;
Compute a histogram of M bins for each θi = {θit}Tt=1, i = 1, …, 5;
Normalize the histograms by dividing T to compensate for different number of frames in a data instance;
Concatenate all normalized histograms into a one-dimensional vector of length 5(M + N );
Convert the feature vector as a single line in the rad d1 or rad d1.t file.
return rad d1 or rad d1.t
1 1 1 2 1 3
0.326 0.325 0.319
-0.101 2.111 -0.058 2.152 0.194 2.166
Each row contains five values, representing:
1) frame id,
2) joint id,
3) joint position x, 4) joint position y, 5) joint position z.
Each frame contains 20 rows that contain information of all joints in the frame.
II. CSCI 473: DELIVERABLE 1 (REPRESENTATION CONSTRUCTION)
Students in CSCI 473 must implement two skeleton-based representations during the Deliverable 1.
A. Relative Distances and Angles of Star Skeleton
Students in CSCI 473 are required to implement the human representation based on the Relative Angles and Distances (RAD) of star skeleton, as described by Algorithm 1. The objective is to implement the RAD representation to convert all data instances in the folder Train into a single training file rad d1, each line corresponding the RAD representation of a data instance. Similarly, all instances in the folder Test needs to be converted into a single testing file rad d1.t.
B. Customized Representations
Implement a customized skeleton-based representation by choosing different joints other than the joints selected in the star skeleton. For example, you can change reference joints, select other joints other than body extremities, or compute distances of all joints but ignore the orientation information. Your code is required to output a single training file cust – d1 for all training instances, with each row containing the customized representation of an instance, and a single testing file cust d1.t, similar to the task in Section VI-A.
C. What to Submit
For Deliverable 1, CSCI 473 students are required to submit a single tarball named D1 firstname lastname.tar (or .tar.gz) to the portal named “P3-D1” in Canvas, which must contain the following items:
• A README that provides sufficient instructions needed to compile and execute your code. Your README also needs to document your implementation information, including which joints are used in the RAD representa- tion, how the histograms are computed, and how many bins are used.
• Your code to construct the RAD and customized repre- sentations.
• The generated representation data, including rad d1, rad d1.t, cust d1, and cust d1.t.
Students are allowed to include a local copy of the training and testing sets within the code directory to make your code self-contained.
III. CSCI 573: DELIVERABLE 1 (REPRESENTATION CONSTRUCTION)
CSCI 573 students are required to implement three specific skeleton-based representations for Deliverable 1, including RAD, HJPD, and HOD.
A. Relative Distances and Angles of Star Skeleton
Students in CSCI 573 are required to implement the human representation based on the Relative Angles and Distances (RAD) of star skeleton, as described by Algorithm 1. The objective is to implement the RAD representation to convert all data instances in the folder Train into a single training file rad d1, each line corresponding the RAD representation of a data instance. Similarly, all instances in the folder Test needs to be converted into a single testing file rad d1.t. This required representation is the same as Section VI-A.
B. Histogram of Joint Position Differences (HJPD)
Given the 3D location of a joint (x, y, z) and a reference joint (xc , yc , zc ) in the world coordinate, the joint displace- ment is defined as:
(∆x, ∆y, ∆z) = (x, y, z) − (xc, yc, zc) (1)
The reference joint can be the skeleton centroid or a fixed joint. For each temporal sequence of human skeletons (in a data instance), a histogram is computed for the displacement along each dimension, i.e., ∆x, ∆y, ∆z. Then, the computed histograms are concatenated into a single vector as a feature.
This HJPD representation is similar to the RAD represen- tation, except that it uses all joints and ignores the pairwise

Fig. 2: The full MSR Daily Activity 3D dataset contains sixteen human activities: (1) drink, (2) eat, (3) read book, (4) call cellphone, (5) write on a paper, (6) use laptop, (7) use vacuum cleaner, (8) cheer up, (9) sit still, (10) toss paper, (11) play game, (12) lie down on sofa, (13) walk, (14) play guitar, (15) stand up, (16) sit down.
Fig. 3: Illustration of human representation based on relative distance and angles of star skeleton
angles. Refer to Section 3.3 of reference [1] for more details, which is available online at: https://ieeexplore. ieee.org/document/6836044.
Your code is required to generate a single training file hjpd d1 for all training instances, with each row containing the HJPD representation of an instance, and a single testing file hjpd d1.t.
C. Histogram of Oriented Displacements (HOD)
You need to implement the skeleton-based representation of Histogram of Oriented Displacements (HOD), as intro- duced in Section 3 of reference paper [2], including the technique of Temporal Pyramid. The paper is available pub- licly at: https://www.ijcai.org/Proceedings/ 13/Papers/203.pdf.
Your code is required to generate a single training file hod d1 for all training instances, with each row containing the HOD representation of an instance, and a single testing file hod d1.t.
D. What to Submit
For Deliverable 1, students in CSCI 573 are required to submit a single tarball, named D1 firstname lastname.tar (or .tar.gz) to the Canvas portal named “P3-D1”, which must contain the following items:
• A README that provides sufficient instructions needed to compile and execute your code. Your README also needs to document your implementation information, for example, including which joints are used in the RAD representation, and how the histograms are computed and how many bins are used in your HJPD and HOD representations.
• All your code to construct the RAD, HJPD, and HOD representations.
• All the generated skeleton-based representation data, in- cluding rad d1, rad d1.t, hjpd d1, hjpd d1.t, hod d1, and hod d1.t.
Students are allowed to include a local copy of the training and testing sets within the code directory to make your code- self contained.
IV. SUPPORT VECTOR MACHINES (PART OF DELIVERABLE 2)
The second deliverable of this project (in Deliverable 2) is to understand and apply Support Vector Machines (SVM) to enable robot learning in practical applications (i.e., behavior understanding in our project).
We will apply LIBSVM as our learning algorithm, which is an excellent open source implementation of SVMs devel- oped by Chang and Lin [3]. The LIBSVM library provides

software support for a variety of SVMs, with the source code available in C++, which also provides an interface to Python and many other programming languages and environments. Here’s the link to the LIBSVM webpage:
http://www.csie.ntu.edu.tw/ ̃cjlin/ libsvm.
Before working on Deliverable 2, you will need to install LIBSVM, get familiar with this library, and understand how to convert the output data files from your Deliverable 1 to the format required by LIBSVM, as described in the following subsections.
A. Installing LIBSVM
First, follow the instructions on the LIBSVM website for downloading the software. The most recent release is Version 3.25 (released on April 14, 2021). Note that the README file within the package provides helpful information for using the LIBSVM package. In addition, you are required to read through and make sure to have a good understanding of the Practical Guide provided by the authors:
http://www.csie.ntu.edu.tw/ ̃cjlin/ papers/guide/guide.pdf.
B. Getting Familiar with LIBSVM
You are required to follow the examples in Appendix A of the Practical Guide to get familiar with how to use LIBSVM. The exemplary datasets (e.g., svmguide1) are available online from the LIBSVM directory, already formatted into the form expected by LIBSVM, here:
http://www.csie.ntu.edu.tw/ ̃cjlin/ libsvmtools/datasets/.
C. Input Data Format Required by LIBSVM
Now, given the histograms, e.g., the skeleton-based repre- sentations such as rad d1 or rad d1.t you have computed in Deliverable 1, you will need to convert them into data files with the a format that can be used by LIBSVM: rad d2 and rad d2.t for training and testing respectively. Specifically, this format of data file now becomes:

1) Read and understand Section IV as well as the Practical Guide of LIBSVM (the theory and practical usage of LIBSVM have been discussed in the class). Go through the examples in the Practical Guide, and understand how to integrate LIBSVM APIs into your source code.
2) Convert the training and testing files (i.e., the outputs of your code to build the representations in Deliverable 1 to a format that can be used by LIBSVM. This must be done for all three representations (i.e., RAD, HJPD, and HOD).
3) Apply LIBSVM to learn a C-SVM model with the RBF kernel from the training data, and use the learned model to predict behavior labels of the testing data, which will generate a result file for all the representa- tions.
4) Write an integrated program that reads data from the training and testing directories, creates a given representation (specified by a command-line flag), per- forms robot learning, and outputs the information of accuracy and confusion matrix to the screen. Your implementation must be based on the LIBSVM APIs (e.g., C++ member functions), and should NOT use the train and test binaries (executables) compiled in the LIBSVM library. Your implementation’s accuracy based on each representation must be better than 60%.
5) Analyze how the accuracy varies according to different numbers of bins.
B. What to Submit
For the Deliverable 2, CSCI 573 students are required to submit a single tarball named D2 firstname lastname.tar (or .tar.gz) to the portal named “P3-D2” in Canvas, which must contain the following items:
• The graphs of the grid search (from grid.py) obtained in the experiments for all the representations (i.e., RAD, HJPD, and HOD);
• What are the “best” values of C and γ of your C-SVMs for all the representations, which can be written in the README;
• All the converted representation files and output predic- tion result files;
• A figure showing how accuracy varies according to the number of bins;
• The code you write in Deliverable 2 and the README (with the same requirement as in Deliverable 1).
VII. DELIVERABLE 3 (PROJECT REPORT)
The third deliverable of this project (Deliverable 3) focuses on writing a project report. The write-up of Deliverable 3 will be posted after Deliverable 2 is due.
VIII. GRADING
The total score of this project is 10 points. Your grade will be based on the quality of your project implementation and the documentation of your findings in the report.
30%: The quality of Deliverable 1. You should have a “working” software implementation, which means that your skeleton-based representations are implemented in software, your code runs without crashing, and performs the behavior understanding task.
45%: The quality of Deliverable 2 (with almost the same requirements to the Deliverable 1).
25%: The quality of Deliverable 3, that is – the project report prepared in LATEX using IEEE robotics conference styling and submitted in the pdf format – Figures and graphs should be clear and readable, with axes labeled and captions that describe what each figure or graph illustrates. The content should include all the experi- mental results and discussions mentioned previously in this document.
Students in CSCI573 will be graded more strictly on the quality of the code implementation and paper presentation. The instructor expects a more through analysis of the ex- perimental results, and a good implementation of skeleton- based representations and robot learning methods for human behavior understanding. The paper should have the “look and feel” of a technical conference paper, with logical flow, good grammar, sound arguments, illustrative figures, etc.
REFERENCES
[1] H. Rahmani, A. Mahmood, D. Q. Huynh, and A. Mian, “Real time action recognition using histograms of depth gradients and random de- cision forests,” in IEEE Winter Conference on Applications of Computer Vision (WACV), 2014.
[2] M. A. Gowayyed, M. Torki, M. E. Hussein, and M. El-Saban, “His- togram of oriented displacements (hod): Describing trajectories of human joints fo

程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com