Algorithm Study On Object Tracking Via Convolutional Neural Network And Network In Network

Object tracking algorithms have gained a great deal of attention in the computer vision community over the past decade due to their potential value both in theoretical challenges as well as in practical applications, including intelligence video surveillance, self-driving vehicles, human computer interaction and so on. Given the initialized object states, the task of object tracking algorithms is to estimate the incoming object states in subsequent frames. Despite much progress has been made in recent years, designing robust object tracking algorithms is still a challenging problem due to object appearance variations caused by pose changes, illumination changes, occlusions, cluttered scenes, moving backgrounds, etc.The limitation of the traditional object tracking algorithms is that the object appearance model is built upon low-level hand-crafted features or shallow classifier,in consequence they have limited predictability power of data semantics. To effectively fill the semantic gap of visual data in object tracking with limited data,inspired by the recent success of deep learning on image classification and speech recognition,we propose novel tracking algorithms which construct a robust object appearance model via learning and transferring mid-level image representations using a deep network structure. Main contents are as follows:1)We propose an object tracking algorithm that relies on convolutional neural network. Traditional object tracking algorithms rely on hand-crafted features which are incapable to capture semantic information of target and not robust to significant appearance changes. Different from hand-crafted features, our algorithm automatically learns the most discriminative features in a data-driven way,and has predictability power of data semantics. Our algorithm is robust to significant appearance changes due to the semantic information and the drifting problem is greatly alleviated. Our algorithm efficiently incorporates the convolutional neural network based appearance model into the particle filtering framework. Experiment results show that our algorithm gets a good performance.2)We propose an object tracking algorithm that relies on network in network.The network in network is used to construct the object appearance model due to its powerful capacity for automatically learning a hierarchical feature representation. In our algorithm, object tracking is formulated as an online transfer learning problem.First, the network in network is pre-trained on the source task. Then, the pre-trained parameters of the internal layers of the network in network are then transferred to the tracking task. This simple yet effective transferring schema enables the proposed tracking algorithm to tackle the domain changes in training tasks. Furthermore, to alleviate the drifting problem, we exploit both the initial and online samples to update the object appearance models. Experiment show that our algorithm achieves more accurate tracking results.