Deep Learning in Gait Recognition for Drone Surveillance System

11th Curtin University Technology, Science and Engineering (CUTSE) International Conference

工业技术（总论）

Sien, Jonathan Phang Then^1 ; Lim, King Hann^1 ; Au, Pek-Ing^1

CDT 250, Sarawak, Miri

98009, Malaysia^1

关键词: Action recognition;    Convolutional neural network;    Deep architectures;    Mobile surveillance systems;    Online computing;    Surveillance systems;    Temporal information;    Video surveillance systems;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/495/1/012031/pdf
DOI  :  10.1088/1757-899X/495/1/012031

学科分类：工业工程学

来源: IOP

【 摘 要 】

As implementation of video surveillance system becomes mainstream, tremendous amount of data is produced. A robust and efficient recognition measure is required for enforcement of public safety in smart city projects. One of a demanded bio-metric recognition measures is gait recognition. With gait recognition, automatic indexing of individual's action in surveillance area allow identification of abnormal activities, hence keeping the area in order. Besides, with advancement in online computing, gait recognition is enabled in remote computing that promotes hybrid of static-mobile surveillance system. Essentially in this research project, a deep learning pipeline to detect suspicious human actions is proposed to recognize individual human's gait via untrimmed continuous streaming video. It consists of Single Shot Multi-box Detector (SSD) for human detection, Inception-V3 based transferred learning convolutional neural networks (CNNs) to extract spatial features. Subsequently, the spatial features are integrated with temporal information by sequentially inserted into Long Short-Term Memory (LSTM) deep architecture for action recognition. The pipeline is trained with KTH human action dataset consists of six action classes, i.e. walking, running, hand clapping, hand waving and boxing. The proposed pipeline achieved 99.51% detection rate to spot the suspicious gait action for surveillance system.