Average Consensus-Based Data Fusion in Networked Sensor Systems for Target Tracking Md Ali Azam Electrical Engineering South Dakota School of Mines and Technology mdali.azam@mines.sdsmt.edu Shawon Dey Electrical Engineering South Dakota School of Mines and Technology shawon.dey@mines.sdsmt.edu Hans D. Mittelmann School of Mathematics and Statistical Sciences Arizona State University mittelmann@asu.edu Shankarachary Ragi Electrical Engineering South Dakota School of Mines and Technology Shankarachary.Ragi@sdsmt.edu Abstract—Decentralized and distributed autonomous sensing over networked sensor systems has many applications in surveil- lance, Internet of Things (IoT), autonomous cars, and UAV swarms tactics. In this study, we develop an average consensus- based decentralized data fusion approach for a target tracking application. Specifically, we extend the standard average consen- sus algorithm to merge the local state estimate information with that of the neighbors. We test the performance of our consensus based data fusion approach for various network configurations. We also perform numerical studies to compare the performance of our approach against the standard Bayesian data fusion approach. Index Terms—Networked sensor systems, Decentralized average consensus, Sensor data fusion, Target tracking I. I NTRODUCTION Autonomous and adaptive sensing has applications such as target tracking, surveillance [1], autonomous car navi- gation [2], and UAV swarm tactics [3], [4]. Particularly, target tracking via adaptive sensing is becoming increasingly important in autonomous car industry for accurate pedes- trian detection and tracking [5]. Sensors such as RADAR, LIDAR, optical sensors, thermal sensors are typically used to measure the target state including its position, velocity, and acceleration. Target tracking with multiple sensors was studied in the past, e.g., [3], where a central fusion node was responsible for making sensing decisions (e.g., sensor location - assuming sensor mounted on a UAV) for all the sensors combined. Clearly, sensing decisions optimized for all the sensors combined provides the best target tracking performance as these decisions are coupled via sensor data fusion. The main drawback with these centralized decision making methods is that they are computationally intensive as the computational complexity is exponential in the decision space and the number of sensors. To address this challenge, we investigated decentralized strategies in the past to some extent [4]. This work was supported in part by Air Force Office of Scientific Research under grant FA9550-19-1-0070. In this study, we develop a decentralized autonomous sens- ing method over a networked sensor system for a target track- ing application. Specifically, we extend an existing approach called average consensus algorithm to perform decentralized data fusion while tracking a moving target. The sensor network is modeled by an undirected graph, which is assumed to be non-time varying. Each sensor generates a noisy measurement of the target state. The presence of an edge between the nodes or sensors means that the sensors are allowed to exchange information/messages for data fusion. In this study, we assume that each sensor maintains a local tracker (or tracking algo- rithm, e.g., Kalman filter), which updates its local target state estimate using the locally generated sensor measurements and the information it receives from its neighbors. We measure the performance of the above consensus algorithm with average target tracking error - the mean-squared error between the target state (ground truth) and the estimate. As a benchmark, we also implement the standard Bayesian data fusion approach for performance comparison. The authors of [6] have surveyed both classical approaches and recent advances in multi-sensor data fusion and consensus filter for sensor networks. The authors of [7] reviewed the key theories and methodologies of distributed multi-sensor data fusion and discussed their advantages like graceful degra- dation, scalability, and interchangeability. Average consensus was studied previously in distributed computing [8] and for achieving consensus among agent values (a real number pos- sibly representing its opinion or state). In [9], a distributed consensus algorithm was developed for obtaining the averages of the node data over networks with large volume of data. N. Gupta et. al. proposed an asynchronous distributed average consensus algorithm [10] to guarantee information-theoretic privacy in multi-agent systems. In [11], the authors provide a theoretical framework for analysis of consensus algorithms for multi-agent networked systems. In [12], the authors developed a distributed consensus tracking filter to solve the target tracking problem. The authors in [13] discussed algorithms for solving decentralized consensus optimization problems. 0964