LEE- Mekatronik Mühendisliği-Yüksek Lisans
Bu koleksiyon için kalıcı URI
Gözat
Son Başvurular
1 - 5 / 30
-
ÖgeReal time pedestrian tracking using adaptive kalman filter(Graduate School, 2022)Image processing has been one of the hot topics since early 1950's and it has developed from simple researches on the images to real time video processing. This development brought more challenges to the researches as the environment around became a complex variable to process. Thus new techniques with low computing time and flexible algorithms that adapt to changes in the environment emerged. Human detection and estimation of their movements is one of the research areas described before. For security and emergency situations, applications of real time human detection is fundamental. In this work it was aimed to develop a robust system that can predict human movement on real time video. The motivation behind the work is to design a system that can detect pedestrians and their movement to warn the driver beforehand. The system needed to do the lowest amount of computation time as possible and it needed to adapt to changes of the environment in order to work in real time with unstable surroundings. Thus an Adaptive Kalman algorithm is developed for prediction of the next steps as only the previous steps information was needed. For the development of the project Matlab is used as the programming language. The reason behind choosing Matlab is due to the fact that Matlab IDE includes many libraries and toolboxes useful for this work. As the first step of the project the camera input is evaluated in real time by taking snapshots and further algorithms are applied on the frames taken. For human detection the built in Matlab library which gets data from the Caltech Dataset is used. The data gathered from the human detection algorithm is used to find the centroids in the region of interest. Centroids are fed to the Adaptive Kalman Filter as the input data with corresponding error parameters. The term Adaptive Kalman Filter comes from the responsive error tuning from the previous input. P, Q and R parameters are tuned with the error values of the previous frame. It was observed that all P, Q and R parameters and K value is converged with the increasing number of frames. Thus the learning process is also included in the work. In the 1000 test frames the system is tested with single and multiple pedestrians in the frame. Out of total 1678 samples 1418 of them are the right classifications which results in a %84.5 success rate. In the test process, it was observed that the error is mainly caused when two objects move towards each other. From that observation it can be concluded that the system works better in areas where the crowd is less dense.
-
ÖgeQuadcopter trajectory tracking control using reinforcement learning(Institute of Science and Technology, 2019-06-11)Unmanned aerial vehicles (UAVs) have gained enormous popularity since the last couple of decades. Quadcopters are the most popular subdivisions of UAVs. Their vertically taking-off, landing and hovering abilities make them ideal platforms for military, agriculture, surveillance and exploration missions. Their mechanical simplicity and agile maneuverability are other two reasons why the quadcopters are so popular. These mentioned reasons make the quadcopters excellent proving grounds for control theory applications. Even though designing a conventional controller for quadcopters is a relatively easy task, tuning those control parameters might easily become a time consuming challenge. Moreover, this requires a model of the system and any uncertainties in the system model or later modifications on the vehicle can quickly cause instabilities. Reinforcement learning is a subclass of artificial intelligence. The idea behind reinforcement learning is making an agent learn in an interactive environment by trial and error principle to achieve a specific task. Notwithstanding it has been discovered long ago, it has got its popularity back with the last advancements in the technology. In this thesis, at first, a conventional PD controller performance on a quadcopter model that is modeled on ETHZ Rotors framework in the Gazebo simulation environment was improved by implementing metaheuristic particle swarm optimization (PSO) algorithm. Thereafter using an actor-critic reinforcement learning algorithm called deep deterministic policy gradient (DDPG), quadcopter was trained to follow different trajectories. DDPG is an off-policy and model-free method, which has proven itself in different domains and tasks. DDPG has four neural network function approximators. These are actor, critic, target actor and target critic networks. The critic network approximates the current value of the agent state and the actor network generates actions with respect to state of the agent. During training, network values shift constantly. Using a constantly shifting set of values to adjust network parameters is not a reasonable thing to do. This makes the value estimations unmanageable. In order to avoid this, DDPG algorithm uses target networks that are used to make the training process more stable. These target networks are not updated at each step, contrary only periodically or slowly updated. Weight decay and batch normalization techniques that are normally not part of the original DDPG algorithm were also implemented to improve algorithm's performance. ADAM algorithm was used for optimization purpose. While training continues, the agent was presented a reward for each step in all episodes. Reward function is defined as negative weighted sum of quadcopter's position, velocity and acceleration errors. Tracking was assumed to be successful, if the tracking error is less than 10%. Tracking performances of both controllers were analyzed for different trajectories. PD controller outperforms reinforcement learning agent in most cases. However, it is needed to be stated that performance differences between two controllers are hard to notice and generalization, which is working on different quadcopter models under some assumptions, is the real advantage of reinforcement learning agent. Hyperparameters of the DDPG algorithm shape the learning behavior of the agent. It is highly possible for a reinforcement learning agent to perform equally or better compared to the conventional controllers. Therefore, as future work, with a given sufficient time, optimizing learning algorithm's hyperparameters and modifying network architectures are worth to investigate in order to have better performances.
-
ÖgeAutomatic landing with model predictive control(Graduate School, 2022)In flight control systems, the landing maneuver is one of the most critical time periods for the aircraft, and it is of great importance to both respond to the disruptors and ensure durability in this time period when the disruptor activity is high. Within the scope of this thesis, 4 different controller type automatic landing systems were designed for a twin-engine passenger aircraft, this landing system provides fully automatic landing in both longitudinal and lateral planes. Within the scope of the thesis, different control architectures in the literature for the automatic landing autopilot were examined. Within the scope of the thesis, the change of system inputs and system outputs as a result of linearization under different conditions has been examined. The consistency of the nonlinear model of the aircraft with the linear model was compared, and this comparison was made by examining the behavior of the system variables in response to the binary commands given to the control surfaces. Within the scope of the thesis, what the sub-phases of the automatic landing autopilot are and according to which criteria and conditions these sub-phases are separated from each other are examined. The classical control architectures in the flight control system (stability-enhancing system and control-enhancing system) are discussed, and for what purposes and with what standards these architectures are designed. In fixed-wing aircraft control systems, the longitudinal and lateral states of the system are separable. In the scope of the thesis, the automatic landing architectures in the literature for these separated states are examined. The controllers designed for the descent system in the thesis are: PID, Linear Quadratic Integral, Model Predictive Control and Algebraic Model Predictive Control architectures. One of these four different control architectures (PID) is in a single-input-single-output control structure, while the other three control architectures (LQI, MPC, AMPC) are in a multi-input-multi-output control structure.
-
ÖgeScalable mechanical design for quadruped robots(Graduate School, 2024-07-10)Nowadays, the vital role of robots in human life is not only undeniable, but it is also essential. Quadruped robots, in particular, which mimic four-legged animals, have been significantly crucial in emergency and critical situations. In this thesis, we deeply focused on different mechanical parameters affecting the design of quadruped robots while considering the dimensional scaling of both robot parts and trajectory length and height, thereby potentially leading us to achieve a scalable control architecture for Quadruped Robots (QRs). The scalability of QRs can significantly enhance their capabilities. If a QR can adjust its size, it can quickly conceal itself under debris to observe enemy operations or navigate through narrow pathways under earthquake rubble. Achieving this ability necessitates a scalable mechanical design, which in turn requires a scalable control architecture. Such an architecture relies heavily on mechanical parameters. To realize this scalable control architecture, it is imperative to meticulously monitor the behavior of these parameters and establish relationships among them. In this thesis, we conducted simulations involving a standard commercial quadruped robot, Unitree A1, walking on a flat surface at a constant speed across five distinct scenarios. Given that contemporary quadruped robots typically feature four motors in the hips and four in the elbows, our simulation followed this configuration, employing a total of eight motors. We just focused on walking forward direction. So lateral movements and turning and other disturbance rejection capabilities are ignored. There are four different sizes of the robot, each expanding the robot size by 30%. In the first scenario, we define the length and height of the trajectory as 120 mm and 27 mm, respectively to observe the effect of scaling. In the second scenario, we increased the length of the trajectory from 120 mm to 165 mm to observe the effect of length increase in trajectory. Then in the third scenario, we defined the length of the trajectory as 165 mm while increasing the height of the trajectory from 27 mm to 40 mm to observe the effect of the step height increase. The fourth scenario maintains the same trajectory length and height as the second scenario, and robot scaling remains consistent except for the Torso to observe the effect of Torso. The Torso retains the dimensions of the Torso of the fourth robot in scenario two (the longest). The linear velocity in all scenarios is 250 mm/s and the robots walk on a flat terrain. Due to the unavailability of precise dimension drawings of the Unitree A1 robot, we endeavored to design its various components in CATIA software, approximating existing robot dimensions. Subsequently, the parts were assembled using SolidWorks software. Leveraging the motion analysis tool within SolidWorks, our thesis aims to generate diverse outcomes, including motor torque, power consumption of the motors, reaction forces, motor angular displacements, linear velocity of the robots, and the mechanical cost of transport (MCOT). By comparing these outcomes, our goal is to establish logical relationships among the mechanical parameters of a standard commercial quadruped robot. The findings of this study hold implications for various actuation design architectures, such as Quasi Direct Drive (QQDs) and series elastic robots. They provide valuable insights that can inform the development and optimization of such architectures while scaling. In conclusion, to the best of our knowledge, there have been no studies examining changes in the mechanical principles of quadruped robots during scaling. Quadruped robots are highly effective in specialized tasks, especially in disaster scenarios like earthquakes, where their mobility outperforms fixed robots. However, altering the dimensions of these robots significantly affects their mechanical and control requirements. This study examined key mechanical parameters, including hip and calf torque, power consumption, reaction forces, and mechanical cost of transport across five different scaling scenarios using simulations with a Unitree A1 quadruped robot. The simulations revealed that while the behavioral patterns of the robots remained consistent, the mechanical demands increased with the elongation of the torso, arms, and legs. Significant changes in angular velocity and displacement of limbs were observed, correlating with motor performance. Successful scaling depends on the motors' ability to handle maximum torque and power consumption requirements while maintaining necessary angular velocity. The study found a consistent mechanical cost of transport (MCOT) across scenarios, with a decrease as trajectory length and height increased, highlighting the importance of minor mechanical variations on energy efficiency. These results provide valuable insights for designing various actuator architectures, not limited to a single actuator type, thereby enhancing their applicability. The research identifies a clear pattern of torques, power consumptions, and reaction forces as the robots scale in size. Future research aims to use this data to develop a scalable control architecture, integrating machine learning. Our research elucidates the behavior of these mechanical parameters during scaling, thereby offering a novel perspective on scalable control architecture in quadruped robots. On the other hand, in scenario five, only the Torso is scaled while the arms and legs retain the dimensions of robot three in scenario two.
-
ÖgeIntegrating path planning and image processing with UAVs for disease detection and yield estimation in indoor agriculture(Graduate School, 2024-07-18)The integration of Unmanned Aerial Vehicles (UAVs) with Controlled Environment Agriculture (CEA) is examined in this thesis, with a particular emphasis on tackling the major obstacles associated with disease detection and yield estimation in indoor farming. In order to maximize indoor agricultural practices, the research intends to take advantage of the advanced capabilities of UAVs fitted with high-resolution sensors together with complex path planning and image processing algorithms. One of the main components of the system is the path planning module. The main task of the path planning module is to effectively guide UAVs through the small areas of indoor farms. This entails flying in the shortest path possible, avoiding obstructions like plant beds, walls and support beams, and making sure the entire greenhouse is covered. Similar to the Traveling Salesman Problem (TSP), the problem is formulated using graph theory and the objective is to find the shortest route that visits all important points. Numerous algorithms, including heuristic approaches like Christofides' Algorithm and Nearest Neighbor, exact methods like Branch and Bound, and metaheuristic strategies like Genetic Algorithms, Ant Colony Optimization, and Simulated Annealing, were assessed. The study came to the conclusion that it would be better to use a combination of heuristic and metaheuristic strategies rather than exact algorithms. They provided the optimum compromise between computational efficiency and solution quality, and they were implemented using Google's OR-Tools library. The path planning module was implemented by generating grid points by using the greenhouse layout and computing distance matrices. Later, the TSP was resolved by refining the early solutions using local search metaheuristics and a variety of first solution methodologies. Path Cheapest Arc, the approach that was selected for the metaheuristic comparisons, showed a consistent rate of path creation, which qualified it for more comparisons and practical deployment. Identifying and counting fruits in snapshots that the UAVs took was part of the yield estimation task in the image processing module. Yolov8, a single-stage detector, was chosen because of its ability to merge speed and precision, which makes it perfect for real-time applications. With a high precision and recall metrics, the YOLOv8s model was trained on a dataset of 8,479 photos that included six different fruit classes. A number of measures were used to assess the model's performance, and the results showed that it was robust and effectively learned, including the Precision-Confidence Curve, F1-Confidence Curve, Recall-Confidence Curve, and Precision-Recall Curve. The main goal of disease detection was to categorize plant leaf images in order to recognize disease symptoms. Latest architectures with great accuracy and computational efficiency, such as YOLOv8s-cls, were selected. A dataset of 18,741 photos, containing both healthy and unhealthy apple and grape leaves, was used to train the model. Confusion matrices and training loss graphs were used to evaluate the model's performance, and they verified the model's dependability and capacity to discriminate between various disease classifications and health states. The ROS and Gazebo platforms were used for system integration and simulation. The UAV platform included key sensors and control algorithms that were integrated with the virtual environment. It was based on the Kopterworx Eagle quadcopter. With this configuration, the control techniques may be thoroughly tested and improved without the hazards that come with actual flight operations. The ROS framework enabled smooth communication between the path planning and image processing components, facilitating modular and distributed system development. The Image Processing node provided real-time picture analysis and precise yield estimation and disease detection while the Path Planner node created effective flight pathways. The UAV was able to function as it would have in a real greenhouse given that to the simulation setup in Gazebo, which imitated a realistic indoor agricultural environment. Throughout the interior setup, the UAV moved steadily and smoothly, accurately following the created flight routes. Real-time processing of the UAV's camera's acquired visuals translated into annotated images that validated accurate yield estimations and accurate disease symptom identification. Through these simulations, the system's capacity to identify unhealthy plants and calculate yields was verified; despite a couple of discrepancies, it demonstrated great accuracy and dependability. The study's findings suggested that an integrated system of unmanned aerial vehicles equipped with innovative path planning and image processing algorithms could substantially improve indoor agriculture's sustainability and efficiency. The dynamic time limit function of the path planning module was essential in guaranteeing effective functioning in different greenhouse sizes. The complexity of the greenhouse arrangement and the quantity of grid points were taken into account by this function while adjusting the time limit dynamically. Through iteratively executing randomized tests for varying point counts, the function determined the point at which solution distances plateud. By minimizing needless delays for simpler layouts and giving enough computing time for complex instances, this adaptive technique allowed the system to maintain superior performance. In the meantime, the image processing module's strong performance indicators highlighted how well it worked in real-time applications. Reducing the dangers and costs associated with physical trials, scalable and cost-effective testing was made possible by the use of the ROS and Gazebo simulation platforms. Additionally, the fruit detection model, tested with real-world images, demonstrated robust performance by utilizing average color analysis to filter grape varieties and reduce false positives, even under challenging conditions. The disease classification model accurately detected and classified leaves, with expert validation recommended to confirm the results, especially under non-ideal conditions. In conclusion, this study showed how combining UAVs with novel technologies might help indoor agriculture overcome its problems with yield estimation and disease detection. In simulated situations, the suggested system demonstrated successful outcomes, demonstrating its capacity to optimize resource allocation, enhance crop management, and guarantee a steady supply of food. Future research should concentrate on scalability and implementation in real-world scenarios, field testing in various indoor farming configurations, and investigating the integration of more sophisticated sensors and enhanced UAV flight dynamics. These developments will improve the system's overall performance in revolutionizing indoor agriculture methods as well as its resilience and adaptability.