DESIGN AND IMPLEMENTATION OF YOLO BASED OBJECT DETECTION AND HYBRID VS-APF BASED QUADROTOR FORMATION CONTROL TO IMPROVE ILLEGAL FISHING MONITORING EFFICIENCY

• Main Author: Halim, Murphy
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/61604
• Institution: Institut Teknologi Bandung

Indonesia has the world's second-largest potential for fisheries. According to President Joko Widodo, the losses from illegal fishing have reached 300 trillion, which is a significant amount when compared to the profits of 65 trillion. Monitoring is carried out around areas suspected of illegal fishing to reduce the number of losses. The monitoring, on the other hand, is ineffective. Monitoring is done with manned aircraft, but the coverage area is still small because it only uses an aircraft, and object recognition in the suspected area is still done manually. As a result, monitoring becomes inefficient in terms of both cost and effort. This study presents a design method for improving the efficiency of illegal fishing monitoring through the use of an unmanned quadrotor formation and object recognition automation. Because the unmanned quadrotor formation requires precision in maintaining its shape, it is designed with formation controllers and path planning based on hybrid algorithms VS (Virtual Structure) and APF (Artificial Potential Field). The design's implementation results in an unmanned quadrotor formation capable of avoiding obstacles by maintaining the formation between quadrotors and achieving the goal position. The controller can maintain the distance between quadrotors with an average absolute error of 11 to 19 % for a 1,5 m setpoint. The YOLOv4 algorithm (You Only Look Once version 4) is used in the object recognition automation design. The YOLOv4 algorithm can produce high-accuracy results in a short amount of time. This design has been demonstrated to be capable of performing object recognition with a mAP (mean Average Precision) of more than 90% and a detection time of less than 26 ms/frame. Overall, the system (formation control, path planning, and object detection) can be successfully integrated and implemented with an accuracy of 83% on a map without obstacles and an accuracy of 77% on maps with obstacles.