Training configuration analysis of a convolutional neural network object tracker for night surveillance application

Abstract

Automated surveillance during the night is important as it is the period when crimes usually happened. By providing continuous monitoring, coupled with a real-time alert system, appropriate action can be taken immediately if a crime is detected. However, low lighting conditions during the night can degrade the quality of surveillance videos, where the captured images will have low contrast and less discriminative features. Consequently, these factors contribute to the problem of bad appearance representation of the object of interest in the tracking algorithm. Thus, a convolutional neural network-based object tracker for night surveillance is proposed by exploiting the deep feature strength in representing object features spatially and semantically. The proposed convolutional network consists of six layers that consist of three convolutional neural networks (CNN) and three fully connected (FC) layers. The network will be trained by using a binary classifier approach of objects and its background classes, which is updated on a fixed interval so that it fully encapsulates the changes in object appearance as it moves in the scene.  The algorithm has been tested with different sets of training data configurations to find the best optimum ones with regards to VOT2015 evaluation protocols, tested on 14-night surveillance videos. The results show that the configuration of a total of 250 training samples with a sample ratio of 4:1 between positive and negative data delivers the best performance for the sequence length of [1,550]. It can be inferred that more information on the object is required compared to the background, where the background might be homogeneous due to low lighting conditions. In conclusion, this algorithm is suitable to be implemented for night surveillance application.