Multi-camera object tracking using surprisal observations in visual sensor networks

Abstract

In this work, we propose a multi-camera object tracking method with surprisal observations based on the cubature information filter in visual sensor networks. In multi-camera object tracking approaches, multiple cameras observe an object and exchange the object's local information with each other to compute the global state of the object. The information exchange among the cameras suffers from certain bandwidth and energy constraints. Thus, allowing only a desired number of cameras with the most informative observations to participate in the information exchange is an efficient way to meet the stringent requirements of bandwidth and energy. In this paper, the concept of surprisal is used to calculate the amount of information associated with the observations of each camera. Furthermore, a surprisal selection mechanism is proposed to facilitate the cameras to take independent decision on whether their observations are informative or not. If the observations are informative, the cameras calculate the local information vector and matrix based on the cubature information filter and transmit them to the fusion center. These cameras are called as surprisal cameras. The fusion center computes the global state of the object by fusing the local information from the surprisal cameras. Moreover, the proposed scheme also ensures that on average, only a desired number of cameras participate in the information exchange. The proposed method shows a significant improvement in tracking accuracy over the multi-camera object tracking with randomly selected or fixed cameras for the same number of average transmissions to the fusion center.

An Effective and Robust Decentralized Target Tracking Scheme in Wireless Camera Sensor Networks

“…${\nabla }_{k-1}={\left[{\ddot{x}}_{k-1}{\ddot{y}}_{k-1}\right]}^T$ is considered as the noise vector, modeled by the zero-mean independent and identically distributed (IID) white Gaussian with covariance matrix ${\mathbf{Q}}_a=diag(\stackrel{false˜}{q_x},\stackrel{false˜}{q_y})$. Subsequently, the motion model of target can further be written as [7] ${\mathbf{x}}_k=f\left({\mathrm{boldx}}_{k-1}\right)+{\mathbf{w}}_{k-1}=\mathrm{boldA}{\mathbf{x}}_{k-1}+{\mathbf{w}}_{k-1},$ where ${\mathrm{boldw}}_{k-1}$ is the process noise vector at timestep $k-1$, assumed the IID white Gaussian with covariance matrix $\mathrm{boldQ}=\left(\begin{array}{cccc}{\stackrel{false˜}{q_x}\delta }^4/4& {\stackrel{false˜}{q_x}\delta }^3/2& 0& 0\\ {\stackrel{false˜}{q_x}\delta }^3/2& {\widetilde{q_x}\delta }^2& 0& 0\\ 0& 0& {\stackrel{false˜}{q_y}\delta }^4/4& {\stackrel{false˜}{q_y}\delta }^3/2\\ 0& 0& {\stackrel{false˜}{q_y}\delta }^3/2& {\widetilde{q_y}\delta }^2\end{array}\right),$ and $\mathbf{A}$ is the state transition matrix: …”

“…Different metrics have been proposed to gauge the tracking performance of the information filter [7,14,22,23]. Among them, the trace (sum of diagonal elements) of the predicted information matrix $Y_{k+1}$ computed at timestep k using Equation (33) corresponds to the mean squared error (MSE) of the updated state.…”

“…In recent years, with the price of smart camera dropping rapidly, the development of wireless camera sensor networks (WCNs) has been heavily fostered [5,6]. Hence, a new trend in target tracking is to deploy sensor nodes with smart cameras to capture, process and analyze image data locally and to send extracted information back to the sink node [7]. However, the target tracking in WCNs is greatly different from that in traditional WSNs with respect to camera field of view (FOV), bandwidth consumption, and multimedia data processing [8].…”

“…In practical circumstances, they are hard to directly apply to WCNs due to the directional sensing and limitations in FOV of cameras. In [7], the authors propose a surprisal selection method to facilitate the camera nodes to take independent decision on whether their observations are informative or not, which considers the directional sensing nature of camera nodes. However, this method fixes the fusion center (FC) and requires the knowledge of the total number of camera nodes that view the current target.…”

Distributed object tracking based on information weighted UAV selection with priory objects

“…Information filters are more suitable for multi-sensor (UAV) object tracking compared to the conventional Bayesian filters due to their inherent information fusion mechanism [8]. In our previous work [9] and [10], we have proposed a multi-camera object tracking method based on the cubature information filter (CIF) with fixed cameras. There, it is shown that the CIF achieves better tracking accuracy than the extended information filter (EIF).…”