Cooperative Search with Multiple Quadcopters using Downward Facing Cameras

Abstract

Multiple robots have been extensively used in performing several tasks cooperatively, in a distributed manner. These distributed multi-robotic systems (MRS) or multi-agent systems (MAS) fi nd application in many fi elds such as search and rescue, environment monitoring, surveillance, landmine detection and clearing, etc. Apart from reduced mission time owing to several agents performing the task simultaneously, distributed MRS are also robust to failure of some of the individual robots. Owing to these advantages, simpler design, and lower cost of individual robots, they are increasingly fi nding applications in adverse conditions such as in military applications and disasters such as natural calamities. Searching for survivors in regions affected by a natural calamity in a civilian context, searching for mines or enemy targets in a military context, using sophisticated sensors carried on unmanned vehicles, such as UAVs or UGVs, are some of the very useful problems that need attention of the researchers from the field of multi-robotic systems. Compared to fi xed-wing UAVs, owing to their maneuverability, ease of takeoff and landing, compactness, lower cost, hovering ability, etc., quadcopters are more suitable for such operations. In this thesis, we formulate a multi-agent search strategy using quadcopter UAVs as search agents/vehicles and downward facing cameras mounted on the quadcopters as search agents. Based on practical considerations, we assumed that the search effectiveness of the camera is maximum at the center and degrades away from it, unlike in most work in the literature where it is assumed to be constant over the entire image frame. The lack of information about presence or absence or the targets of interest in the search space is modeled as an uncertainty density distribution. Here, the uncertainty is 1 when no information on the existence (or absence) of the target at a point of interest is available and 0 when it is established that the target is either present or absent at that point. Based on uncertainty density distribution and the monotonically decreasing search effectiveness model, we address and formulate the problem of optimally deploying the quadcopters so as to maximize the uncertainty reduction (and hence information gain). Based on the observation we make on similar problem setting used in the literature, we formulate a `deploy' and `search' strategy using the concepts of centroidal Voronoi confi guration, where the quadcopters get deployed to a centroidal Voronoi con figuration, shown to be an optimal con guration maximizing the reduction in uncertainty, and then perform search resulting in a reduction in the uncertainty. The process of optimal `deployment' and `search' continue until the average uncertainty over the entire search space is reduced below an arbitrary but fi xed value, indicating the targets, if presented, are detected with an acceptable confi dence (probability). One of the very important components in multi-agent search is the search sensor itself and the spatial variation of its effectiveness in performing the search, that is target detection. As we mentioned earlier, we assume the non-uniform effectiveness of the camera within its image frame. We rst provide a detailed discussion on the spatial variation of the image quality both in terms of optical resolution and digital quality. We observe that the image quality is higher at the central pixel and degrades away from it. Such a scenario leads us to a nonuniform search effectiveness of the camera. We present an experimental setup to obtain a sensor effectiveness model for a downward-facing camera using the target detection probability. Through a set of target detection experiments carried out using AuRuco markers and triangular-shaped objects as targets, we obtain a sensor effectiveness model for a downward-facing camera in different scenarios. We also establish that an exponential function with two parameters can be used to model the spatial variation of the camera's search effectiveness (that is, the search effectiveness model). We develop a platform using ROS/Gazebo and Matlab environment for simulation of the proposed multi-quadcopter search strategy in a hybrid centralized-decentralized architecture. The platform developed can be a very useful tool for conducting realistic simulation experiments to validate the proposed search strategy and to make a comparative study of its performance in terms of time required for the search process, with different parameters such as camera search effectiveness functions, sensor range, number of robots, and decide on the right parameters for any given mission. We provide detailed results of experiments and simulation carried out along with a detailed discussion on the same. First, we present the results of the experiments carried out to obtain the search effectiveness model of the downward facing camera, which we use in the proposed multi-quadcopter search strategy. Though we used an experimental setup to establish, that in general, an exponential function with two parameters can be as the search effectiveness model of a camera, it can be used to carry out experiments with a specific c type of imaging sensor, the type of image processing tools, the kind of environment in which it has to detect the targets, the type of targets that need to be detected, and hence obtain a suitable search effectiveness model. We present representative results of the simulation experiment carried out using the realistic ROS/Matlab simulation platform, both to demonstrate the simulation platform itself and the proposed search strategy. Finally, we provide a detailed account of simulation experiments carried out to evaluate the effect of the number of search quadcopters and the camera effectiveness parameters on the performance of the proposed multi-quadcopter search strategy, using the simulation platform developed in this work. The simulation platform developed can be used to carry out experiments using physical AR Drones. The controller used within the simulation environment may be used to control the physical AR Drones. Also, the simulation environment can be used to conduct a large number of simulation and physical experiments to decide on parameters such as the optimal number of quadcopters, type of cameras used (in terms of their search effectiveness, which may be obtained by using the experimental setup based on that used in this work), for a given search scenario. In this sense, the experimental setup and simulation platform developed are useful beyond the sample results provided in this thesis and will surely help the proposed multi-agent search strategy takes a step forward from theory to experiment and then fi nally into reality.