Deep-learning based single object tracker for night surveillance

Abstract

Tracking an object in night surveillance video is a challenging task as the quality of the captured image is normally poor with low brightness and contrast. The task becomes harder for a small object as fewer features are apparent. Traditional approach is based on improving the image quality before tracking is performed. In this paper, a single object tracking algorithm based on deep-learning approach is proposed to exploit its outstanding capability of modelling object’s appearance even during night. The algorithm uses pre-trained convolutional neural networks coupled with fully connected layers, which are trained online during the tracking so that it is able to cater for appearance changes as the object moves around. Various learning hyperparameters for the optimization function, learning rate and ratio of training samples are tested to find optimal setup for tracking in night scenarios. Fourteen night surveillance videos are collected for validation purpose, which are captured from three viewing angles. The results show that the best accuracy is obtained by using Adam optimizer with learning rate of 0.00075 and sampling ratio of 2:1 for positive and negative training data. This algorithm is suitable to be implemented in higher level surveillance applications such as abnormal behavioral recognition.