Robust visual tracking based on convolutional neural network with extreme learning machine

Abstract

Recently, deep learning has attracted substantial attention as a promising solution to many problems in computer vision. Among various deep learning architectures, convolutional neural network (CNN) has demonstrated superior performance as a feature learning method. In this paper, we present a novel hybrid model of CNN and extreme learning machine (ELM) for object tracking. Training a conventional CNN requires a substantial amount of computation and a large dataset. ELM randomly generates the parameters of hidden layers and calculates network weights between output and hidden layers via the regularized least-square method, thereby dramatically reducing the learning time while producing accurate results with minimal training data. Therefore, we integrate the ELM auto-encoder architecture into the CNN model. In addition, an effective updating scheme is designed for the model training to overcome the tracking drift problem. The joint CNN-ELM tracker is robust to object variations such as illumination, occlusion, and rotation in a video sequence. Numerous experiments on various challenging videos demonstrate that the proposed tracker performs favourably compared to several state-of-the-art methods.