IMPLEMENTATION OF PARALLEL NAVIGATION AND MULTIPLE OBJECT TRACKING FOR TARGET DETECTION, PURSUIT, AND OBSTACLE AVOIDANCE USING A DRONE SWARM

• Main Author: Rahardian A S, Reinard
• Format: Final Project
• Language: Indonesia
• Subjects: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• Online Access: https://digilib.itb.ac.id/gdl/view/70289
• Institution: Institut Teknologi Bandung

Drone is a technology that allows convenient object retrieval, delivery, or humanity mission on places where the terrain is too unfavorable for the mission to be done directly by humans. One example of such mission is pursuit, of animals for conservation purposes, or of humans for law enforcement cases. While this does sound promising, usage of drones for these missions are hindered by difficult control and intelligence implementation that fails to approach a human’s flexibility, in addition to perception limitations on sensors mounted on a drone. This research is intended to overcome said limitation by implementing a system that allows pursuit to be done by a swarm of drones, thus solving the intelligence and perception problem by adding pursuers that can help in surrounding the pursuit target and having extra active sensors on the additional drone. The system will be constructed out of a controller system consisting a High-Level Controller based on parallel navigation that allows pursuit against a target while avoiding obstacles, and a Low-Level Controller based on Proportional-Integral- Derivative controller. Additionally, an image processor system will also be implemented to estimate the exact location of objects that might appear around the drone’s surrounding. This image processor will use YOLOX object detector to infer object’s direction of appearance, then triangulate the exact coordinate by using Iteratively Reweighted Midpoint Method (IRMP) Triangulation algorithm. Simulations and experiments have been conducted using an off-the-shelf quadrotor drone Parrot AR Drone 2.0. PID controller modelling has been performed against the drone’s attitude controller of which characteristics has been modeled on previous researches. The resulting controller has a 5% settling time of 1.39, 1.52, 0.97, and 1.32 seconds for x-axis velocity, y-axis velocity, height, and yaw orientation controllers respectively. The High-Level Controller based on parallel navigation has successfully been implemented and simulated to control a swarm of 1 to 4 drones. Results show that a swarm of 4 drones always locate the target within 1.2 seconds and surround the target on average in 11.67 seconds. An image processor has also been implemented using YOLOX-L and IRMP with performance benchmark at 10.3 inferences per second, and testing shows that triangulation performed with 5 or more viewpoint data results in reconstruction error less than 1 meter, while the reconstruction loss with 3 to 24 data is within 5 to 32 pixels. This image processor is then integrated with the controller system in a field test using Parrot AR Drone 2.0. Testing is performed by flying a drone twice, each time as a different member of a swarm of 2 drones, and replaying the flight data record from the first flight on the second flight to simulate simultaneous flight. This testing shows the swarm locates the target within 6.03 seconds on average, and fully surrounds the target on average 8.53 seconds after locating the target.