Real-time deep learning based system to detect suspicious non-verbal gestures
Feba Thankachan George, Venkata Sindhoor Preetham Patnam, Kiran George · 2018
As crime rates are increasing all over the world these days, recognizing it in advance will help to reduce it to a significant level. Various studies show that some common signs of non-verbal gestures along with our intellectual judgment can assist to identify a subject with suspicious intentions successfully. A system that utilizes deep learning techniques to detect suspicious non-verbal gestures in real-time is presented in this paper. The system utilizes Jetson TX1 development kit from NVIDIA® and multiple body cameras to recognize suspicious non-verbal gestures associated with limbs or torso. The architecture used for non-verbal gesture training and classification is Convolution Neural Network (CNN) developed in Caffe, which is a deep learning framework. Training the system to classify the non-verbal gestures resulted in an accuracy of ~93% and an accuracy of ~91% is achieved when real-time functional testing is carried out on four healthy subjects. With additional training, system's recognition accuracy can be improved.