In recent years, quadrotor unmanned aerial vehicles (UAVs) have gained significant traction across various military and civilian applications due to their agility and versatility. However, the ability to achieve fully autonomous flight remains a critical challenge, particularly in dynamic environments where obstacles and precise positioning are paramount. As part of an educational initiative under the CDIO project practice curriculum, our team embarked on designing and implementing a quadrotor system capable of autonomous navigation, obstacle avoidance, and target-reaching tasks. This project focuses on integrating multiple technologies, including GPS, binocular vision, LiDAR, and artificial potential field methods, to enable a quadrotor to plan paths, avoid obstacles, and accurately reach designated target points. The quadrotor’s design emphasizes modularity using the Robot Operating System (ROS) framework, ensuring scalability and real-time performance. Through rigorous testing, we have demonstrated that our quadrotor can successfully navigate complex routes, circumvent obstacles, and complete missions such as material delivery, showcasing the potential of autonomous quadrotor systems in real-world scenarios.
The structural design of the quadrotor is fundamental to its stability and functionality. We selected a robust F450 frame as the core structure, characterized by a cross-shaped, axis-symmetric configuration with a diagonal motor distance of 450 mm. This quadrotor comprises several key components: the frame, onboard computer, motors and propellers, battery, processor, various sensors, and electronic speed controllers (ESCs). The mechanical architecture is divided into six main parts: elevation brackets, middle layer, arms, battery compartment, camera mounts, payload frame, and landing gear. The elevation brackets elevate the LiDAR and safety switch to prevent obstruction of the scanning path and facilitate safe access. The middle layer serves as the primary support, connecting the arms, battery compartment, elevation brackets, and landing gear, while housing the PX4 flight controller, onboard computer, and ESCs. The arms, constructed from aluminum square tubes for lightweight durability, transmit lift from the motors and propellers to the main body. The battery compartment and camera mounts secure the battery and connect binocular cameras and the payload frame, which holds materials and a monocular camera for targeted drops. To mitigate impact during landing, the landing gear is wrapped with shock-absorbing foam. This comprehensive design ensures the quadrotor remains stable and efficient during autonomous operations.
Algorithm design is crucial for enabling the quadrotor to autonomously plan paths, avoid obstacles, and reach multiple target points. Our approach leverages ROS for modular development, dividing the system into key modules: autonomous positioning, visual recognition, obstacle detection, and path planning. The autonomous positioning module combines GPS for large-scale outdoor navigation and binocular vision for precise indoor or short-range localization. Binocular vision relies on the Scale-Invariant Feature Transform (SIFT) algorithm to extract and track feature points across different scales. SIFT identifies key points such as corners, edges, and intensity variations by constructing a Gaussian scale-space. The Gaussian blur process uses a convolution matrix derived from the two-dimensional Gaussian function:
$$ G(x,y) = \frac{1}{2\pi\sigma^2} e^{-\frac{x^2 + y^2}{2\sigma^2}} $$
where $ \sigma $ is the standard deviation, and $ x $ and $ y $ represent the pixel coordinates in the blur kernel. This process creates a Gaussian pyramid by repeatedly blurring and down-sampling the image, allowing feature extraction at multiple scales. For binocular vision, the depth $ z $ of a feature point $ P $ is calculated using triangulation:
$$ z = \frac{f \cdot b}{x_l – x_r} $$
where $ f $ is the focal length, $ b $ is the baseline distance between cameras, and $ x_l $ and $ x_r $ are the x-coordinates of the feature point in the left and right camera images, respectively. By tracking multiple feature points, the quadrotor’s displacement relative to a fixed coordinate system is derived, enabling accurate positioning. This data is transmitted via ROS topics to the coordination control module for further processing.
Visual recognition is employed to identify target points for material delivery, utilizing the YOLOv3 neural network for real-time object detection. We trained the model on a dataset of printed patterns and real-world images, with parameters optimized for efficiency and accuracy. The training configuration is summarized in Table 1.
| Parameter | Value |
|---|---|
| Epochs | 200 |
| Batch Size | 64 |
| Subdivisions | 32 |
| Learning Rate | 0.001 |
The YOLOv3 architecture processes input images through convolutional layers to detect objects and output bounding boxes with confidence scores. This allows the quadrotor to locate target points accurately, with results published as ROS topics for coordination.
Obstacle detection is achieved using LiDAR technology, which provides reliable point cloud data unaffected by environmental conditions. The LiDAR sensor scans a 360-degree plane, measuring distances to obstacles and generating point clouds. We implement Simultaneous Localization and Mapping (SLAM) techniques to construct 2D and 3D maps in real-time. To enhance 3D mapping, noise reduction and coupling with positioning data are applied, ensuring the quadrotor can detect obstacles in its flight path. The point cloud data is converted to ROS LaserScan messages and forwarded to the coordination module for obstacle avoidance.
Path planning and obstacle avoidance are managed through the artificial potential field method, which superimposes attractive forces from targets and repulsive forces from obstacles. The quadrotor moves along the resultant force vector, ensuring it avoids collisions while progressing toward goals. The attractive force $ F_g $ in a given direction is computed as:
$$ F_g = C_1 \cos \theta $$
where $ C_1 $ is the attraction coefficient, and $ \theta $ is the angle between the direction and the target. The repulsive force $ F_r $ is given by:
$$ F_r = \frac{C_2}{d} $$
where $ C_2 $ is the repulsion coefficient, and $ d $ is the distance to the nearest obstacle in that direction. If $ d $ exceeds the LiDAR range, the maximum range value is used. The resultant force $ F $ in each direction is:
$$ F = F_g – F_r $$
The quadrotor evaluates forces at intervals of $ \Delta \theta $ over a 180-degree frontal arc and selects the direction with the highest resultant force for navigation. Parameters for this method are listed in Table 2, where $ F_{\text{min}} $ is the minimum force threshold.
| Parameter | Value |
|---|---|
| $ C_1 $ | 10 |
| $ C_2 $ | 4 |
| $ F_{\text{min}} $ | 0 |
Extensive testing was conducted to validate the algorithms and system integration. For binocular vision, we performed feature point extraction tests using SIFT, ensuring that at least five feature points were tracked continuously during quadrotor movement. In a closed-loop test within a 4 m × 3.5 m × 2 m space, the quadrotor returned to its starting point with positional errors of Δx = 0.12 m, Δy = -0.06 m, and Δz = 0.01 m, demonstrating sufficient accuracy for autonomous tasks. Visual recognition tests confirmed that the YOLOv3 model could rapidly identify target patterns with high confidence, enabling real-time decision-making. For overall system validation, we designed a mission where the quadrotor had to follow a predefined route: start point → deliver material at point A → deliver at point B → deliver at point C → return to start, while avoiding obstacles. Both simulation and real-flight tests showed that the quadrotor successfully navigated the course, avoided obstacles, and completed material drops, highlighting the effectiveness of our integrated approach.
In conclusion, this project has successfully developed a quadrotor system capable of autonomous positioning and obstacle avoidance. By leveraging binocular vision with SIFT feature extraction, YOLOv3 for target recognition, LiDAR-based obstacle detection, and artificial potential field path planning, we have created a robust platform for autonomous quadrotor operations. The modular ROS framework ensures flexibility and scalability, allowing for future enhancements. This work, inspired by the CDIO project practice and national engineering competitions, underscores the potential of quadrotor UAVs in applications such as logistics and surveillance. Future efforts will focus on improving localization accuracy under adverse conditions and expanding the quadrotor’s capabilities for more complex environments.