الفهرس 
Autism, a literature reviewOnly 14 pages are availabe for public view
 |  | from | 181 |  |  |
| | | from | 181 | | |
Abstract
Abstract
Autism Spectrum Disorder (ASD) is a mental developmental disorder associated with social and communicational defects and Stereotypical Motor Movements (SMM). SMM is also associated with several mental developmental disorders and has several forms like arm flapping, head banging, ear covering and spinning with various degrees of severity. SMM affects the learning development of children and might lead to self-injury in severe cases.
ASD reasons are still mysterious, but scientists believe that genetic defect is the main cause. Although autism is a lifelong disorder; early diagnosing and treatment can improve children’s development stages.
Diagnosing autism has been an exhaustive process that attracts several researchers’ attention. The traditional protocol of diagnosing autism relies on long clinical observation sessions for an autistic child. As the population of autistics exceeds the medicals who can diagnose this disorder, the diagnosing waiting time may exceed six months.
Since the traditional protocol of diagnosing autism was a major time-consuming problem, several techniques were utilized to assist in diagnosing autism including Electroencephalography signals (EEG), Functional Magnetic Resonance Imaging (fMRI), Blood and genetics analysis, wearable sensors, and computer vision techniques. However, computer vision-based techniques have many advantages over the other techniques in terms of simplicity and minimized costs.
This thesis aims to design and implement a computer vision-based system that detects and recognizes some signs of autism in children from video sequences, focusing on repetitive motor behaviours in autistics.
Three data sets are used to evaluate the proposed model; data set (A) which is self-stimulatory behaviour dataset SSBD that contains 25 videos of children performing arm flapping,headbanging and spinnin, data set (B) was 20 videos collected from public domains of children performing arm-flapping autism and headbangin symptoms, data set (C) containing recorded videos of well-developed people, ageing from three to twenty-six years, mimicking arm flapping and head-banging. SSBD videos were very challenging due to the nature of the recording circumstances and the quality of the cameras used.
The best accuracy was achieved from the neural network model that consisted of 12 fully connected layers, each layer has 128 neurons, with Mean Squared Error (MSE) loss function and ADAMAX optimization function. Piecewise constant decay function was applied to the learning rate to minimize the fluctuation caused by the mini-batch gradient descent on the model.
This model was trained for 300 epochs producing 99.9% training accuracy, 99.5% validation accuracy and 100% testing accuracy.
To our knowledge, our proposed model produced state of the art accuracy in detecting arm flapping on SSBD and showed its efficiency and reliability on other different real-world videos enabling it to be further used for different self-stimulatory symptoms.
Head Banging represents another daily repetitive motor behaviour characterizing autism. A proposed neural network model is developed for head banging detection and recognition. The proposed model consists of an input layer with the size of features extracted by the OpenPose model, followed by twelve fully connected layers. Each layer consisted of 128 neurons with Tanh activation function. A flatten layer is added to transform the features into a one-dimensional vector. The last layer was the output layer which was activated by SoftMax activation. The used loss function was mean squared error MSE in cooperation with the ADAMAX optimization function. Learning rate decay was applied to the learning rate parameter to decrease the fluctuation in the learning accuracy appeared using mini-batch gradient descent. The proposed solution was trained for 300 epochs resulting in 99.5% training accuracy, 98.24% validation accuracy and 99% for testing accuracy.
Despite the challenges faced during processing the self-stimulatory behaviour data set (SSBD), our model achieved 85.5% accuracy of head banging detection. In addition to that, our proposed model results in better accuracy on the collected data set in which it has better recording circumstances and higher videos quality. Our proposed model achieved 93% accuracy of head banging detection of our collected and videoed data set.
For all we know, our proposed model achieved a state-of-the-art accuracy for head banging behaviours detection using vision-based techniques on SSBD. In addition to that, our model proved its efficiency and reliability by achieving 93% accuracy on our collected and videoed data set. The achieved results imply that repetitive motor behaviours can be accurately detected and recognised accurately using vision-based techniques in cooperation with neural networks.