Professional Documents
Culture Documents
Department of Robotics and Mechatronics Engineering: Project / Dissertation
Department of Robotics and Mechatronics Engineering: Project / Dissertation
RME 410
Project / Dissertation
Deep Model Based Autism Spectrum Disorder (ASD)
Detection through Activity Recognition
Presented by- Supervised by-
Omar-Ibne-Shahid Dr. Sejuti Rahman
Exam Roll: 413 Assistant Professor,
Registration ID: 2015-716-909 Department of Robotics
and Mechatronics Engineering,
University of Dhaka
5 January, 2020
2
Outline:
Introduction
Objectives
Related Works
Dataset Description
Methodology
Comparative Results
Conclusions
3
Introduction(1/8)
Ref.1 “Centers for disease control and prevention.” https://www.cdc.gov/ncbddd/ autism/data.html, 2014.
Ref.2 F. Rahman, S. Akhter, A. Biswas, and A.S. Abdullah, “Study of prevalence of autism in
Bangladesh,” 2016.
Related Dataset Comparative
Introduction Objectives Methodology Conclusion
Works Description Results
6
Introduction(4/8)
MRI Image Analysis
ASD
Detection
Techniques
1. Developmental Screening
Developmental screening is a short test to
tell if children are learning basic skills at ASD
proper time Detection
2. Comprehensive Diagnostic Evaluation Techniques
Analyze child’s behavior, expressions,
interaction and relation with others
Future
Prospects
Automated
System
ASD vs. TD
Related Dataset Comparative
Introduction Objectives Methodology Conclusion
Works Description Results
12
Objectives(2/3)
Applying
Deep
Model
What is
Human activity recognition is
an ability to interpret human Human
body gesture or motion via Activity
sensors and determine human
activity or action Recognition
(HAR)
Ref.3 M. J. Roshtkhari and M. D. Levine, “Human activity recognition in videos using a single example,”
Image and Vision Computing, vol. 31, no. 11, pp. 864–876, 2013
Classification
of HAR
Ref.4 O. Rihawi, D. Merad, and J.-l. Damoiseaux, “3D-AD: 3D-autism dataset for repetitive behaviors with
Kinect sensor,” in 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance
(AVSS), pp. 1–6, IEEE, 2017
Ref.5 Q. Guillon, N. Hadjikhani, S. Baduel, and B. Rog´e, “Visual social attention in autism spectrum disorder:
Insights from eye tracking studies,” Neuroscience & Bio-behavioral Reviews, vol. 42, pp. 279–297, 2014
Ref.6 D. Kindregan, L. Gallagher, and J. Gormley, “Gait deviations in children with autism spectrum disorders:
a review,” Autism research and treatment, 2015
Unavailability of Datasets
Dataset Diversity
Data Annotation
Challenges
Experimental Setup
in this
Capturing Long Context domain
Privacy Issues
Ref.7 A. Zunino, P. Morerio, A. Cavallo, C. Ansuini, J. Podda, F. Battaglia, E. Veneselli, C. Becchio, and V. Murino, “Video
gesture analysis for autism spectrum disorder detection,” in 2018 24th International Conference on Pattern Recognition
(ICPR), pp. 3421–3426, IEEE, 2018
Popular
3D convolutions applies a 3 Models
dimensional filter to the dataset
and the filter moves 3-direction to Classify
(x, y, z) to calculate features Video
Gestures
Popular
Models
VGG-16 to Classify
Video
Gestures
Ref.8 K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image
recognition,” arXiv preprint arXiv:1409.1556, 2014
Related Dataset Comparative
Introduction Objectives Methodology Conclusion
Works Description Results
25
Related Works(12/13)
CNN LSTM Output
Popular
Models
ResNet-50 to Classify
Video
Gestures
Ref.9 K. He, X. Zhang, S. Ren and J. Sun, "Deep Residual Learning for Image Recognition," IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2016
Popular
Models
Inception v3 to Classify
Video
Gestures
Ref.10 C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for
computer vision,” IEEE conference on computer vision and pattern recognition, pp. 2818–2826, 2016.
Foundation
Bottle grasping
Activity
1. grasp-to-place
2. grasp-to-pour
3. grasp-to-pass to place
4. grasp-to-place to pour
Experimental
Setup
Data
Preparation
Original Filtered
4 Action class-
1. grasp-to-pour
2. grasp-to-place Data
3. grasp-to-pass to pour
4. grasp-to-pass to place
Preparation
2 class-
1. ASD
2. TD
VGG-16/
ResNet-50/
Inception v3
Approaches
FC Layer
LSTM
3D CNN
Output
CNN + LSTM
Related Dataset Comparative
Introduction Objectives Methodology Conclusion
Works Description Results
34
Methodology(4/10)
3D
CNN
• Number of epoch: 10
• Batch size: 56 3D
• Loss function:
Categorical Cross-
CNN
entropy
• Optimizer: Adadelta
• Learning rate decay:
0.95
3D
CNN
Transfer
Learning
Transfer
Learning
Transfer
Learning
Simple
3D CNN
43.34 37.18
4 Action
Modified 63.67 58.39 Class
3D CNN
Previous
Work
Previous
Work
Simple 49 44 2 Class
3D CNN
Modified 68 60.35
(ASD & TD)
3D CNN
Accuracy
Model Train Test
(%) (%)
Accuracy
Lower Accuracy
Can’t Predict Autism Level Limitations
Applicable for Experimental Constrained Dataset
Enlarging Dataset
Autism Level Annotation using DSM-5 Screening Future
Train Model from Scratch Scopes
Proper Tuning of Parameters