Convolutional Neural Networks Tracker with Deterministic Sampling for Sudden Fall Detection

Fall-down event detection has increasingly become more important for surveillance systems, which has been catalyzed by widespread availability of surveillance cameras in public places. This event detection plays a vital role in ensuring one’s safety, especially in public health applications such as monitoring system for toddlers, patients, and elderly peoples. Fall-down event detection has been extensively researched because of price decrements in video-based sensor and high-speed computer networks. This paper has adopted a deterministic sampling approach to train the object appearance model using a multiple-model fully convolutional neural network (MMFCNN) tracker. The CNNs tracker comprises of three convolutional layers and three fully connected layers to model the object appearance with minimal user intervention except during the initialization stage. The ground truth location of the object in the first frame is the sole user input information that will be fed to the MMFCNN tracker. The positive training samples are generated by translating the object ground truth location in eight different directions with a step size, k=1. On the other hand, the negative training samples are generated from eight-neighborhood regions of equal size that surround the object in a clockwise rotation direction. The experimental results show that MMFCNN tracker has better performance in term of robustness and reliability, which are crucial for fall down event detection as compared to TCNN and MDNET.