Introduction
Micro-Aerial Vehicles (MAVs) are ideal platforms for indoor and outdoor applications because to their small size and light weight [1, 2]. Obstacles, on the other hand, may cause MAVs to crash. Cameras collect a lot of evidence about their surroundings. Through the use of grayscale values [3], point features [4], and edge details [5], vision-based techniques detect obstacles. They are divided into monocular and stereo types. There are four types of monocular approaches: appearance-based [6], motion-based [7], depth-based [8], and expansion-based [4]. Expansion-based techniques are based on the same principle as human vision. The majority of expansion-based systems identify obstacles by recognizing object points [4, 5, 9-11]. However, relying solely on points may not be sufficient; as a result, the MAV may collide with unseen impediments. To overcome this obstacle, we describe a novel technique based on the same concept but employing region-enlarging rates.
Methodology
Various steps of our obstacle detection technique above can be summarised in (a) data acquisition and preparation, (b) region extraction and their area calculation, and (c) obstacle detection.
Results and discussion
We took four pairs of images with an LG 360 CAM fisheye camera, two with the camera moving forward and two with the camera moving to the sides. In the forward direction, recall accuracy is 82% for the first data and 52% for the second data. The new technique detects only a portion of the obstacle region. This problem emerges because some regions lack at least three matching points. While moving to the right, recall accuracy for the third and fourth data is 69% and 39%, respectively. This accuracy is lower in the fourth data set than in the other data sets due to the aforementioned description, the absence of at least three corresponding points, and the possibility of inaccurate corresponding points, particularly along the fisheye image's edges, which have a low quality in these places.
Conclusion
The proposed method extracts regions of close obstacles from outdoor fisheye images. The findings demonstrate the method's efficacy in a variety of complex environments. Thus, on average, 60% of the obstacles are detected in two modes of forward movement and right movement. Additionally, a comparison of the suggested method to that of Al-Kaff et al. (2017) [4] demonstrates that it is more efficient than the existing algorithm. The proposed algorithm, however, has some gaps in terms of obstacle detection.One of these limitations is that certain regions lack at least three corresponding points. Also, the presence of incorrect corresponding points causes incorrect detection of obstacles. The second limitation is that the obstacle is the same color as the background, which leads to errors in the correct detection of obstacles. The third constraint is the long processing time required by the suggested approach. These constraints can be solved in the future with the use of more accurate and faster algorithms. |
