Vision based Collaborative Path Planning for Micro Aerial Vehicles Conference Paper uri icon

abstract

  • Autonomous micro aerial vehicles (MAV) have gained immense popularity in both the commercial and research worlds over the last few years. Due to their small size and agility, MAVs are considered to have great potential for civil and industrial tasks such as photography, search and rescue, exploration, inspection and surveillance. Autonomy on MAVs usually involves solving the major problems of localization and path planning. While GPS is a popular choice for localization for many MAV platforms today, it suffers from issues such as inaccurate estimation around large structures, and complete unavailability in remote areas/indoor scenarios. From the alternative sensing mechanisms, cameras arise as an attractive choice to be an onboard sensor due to the richness of information captured, along with small size and inexpensiveness. Another consideration that comes into picture for micro aerial vehicles is the fact that these small platforms suffer from inability to fly for long amounts of time or carry heavy payload, scenarios that can be solved by allocating a group, or a swarm of MAVs to perform a task than just one. Collaboration between multiple vehicles allows for better accuracy of estimation, task distribution and mission efficiency. Combining these rationales, this dissertation presents collaborative vision based localization and path planning frameworks. Although these were created as two separate steps, the ideal application would contain both of them as a loosely coupled localization and planning algorithm. A forward-facing monocular camera onboard each MAV is considered as the sole sensor for computing pose estimates. With this minimal setup, this dissertation first investigates methods to perform feature-based localization, with the possibility of fusing two types of localization data: one that is computed onboard each MAV, and the other that comes from relative measurements between the vehicles. Feature based methods were preferred over direct methods for vision because of the relative ease with which tangible data packets can be transferred between vehicles, and because feature data allows for minimal data transfer compared to large images. Inspired by techniques from multiple view geometry and structure from motion, this localization algorithm presents a decentralized full 6-degree of freedom pose estimation method complete with a consistent fusion methodology to obtain robust estimates only at discrete instants, thus not requiring constant communication between vehicles. This method was validated on image data obtained from high fidelity simulations as well as real life MAV tests. These vision based collaborative constraints were also applied to the problem of path planning with a focus on performing uncertainty-aware planning, where the algorithm is responsible for generating not only a valid, collision-free path, but also making sure that this path allows for successful localization throughout. As joint multi-robot planning can be a computationally intractable problem, planning was divided into two steps from a vision-aware perspective. As the first step for improving localization performance is having access to a better map of features, a next-best-multi-view algorithm was developed which can compute the best viewpoints for multiple vehicles that can improve an existing sparse reconstruction. This algorithm contains a cost function containing vision-based heuristics that determines the quality of expected images from any set of viewpoints; which is minimized through an efficient evolutionary strategy known as Covariance Matrix Adaption (CMA-ES) that can handle very high dimensional sample spaces. In the second step, a sampling based planner called Vision-Aware RRT* (VA-RRT*) was developed which includes similar vision heuristics in an information gain based framework in order to drive individual vehicles towards areas that can benefit feature tracking and thus localization. Both steps of the planning framework were tested and validated using results from simulation.

author list (cited authors)

  • Vemprala, S., Saripalli, S., & IEEE, ..

publication date

  • April 2018