Akademik Veri Yönetim Sistemi
Journal of Network Intelligence, cilt.8, sa.2, ss.421-432, 2023 (Scopus)
Obstacle detection is an important research direction in the field of mobile robot avoidance. Traditional binocular recognition mainly uses binocular cameras to obtain the disparity map of each object in the field of view to measure the distance of the object. In this way, the depth information about the obstacle can be well obtained. However, the disparity map does not distinguish obstacles and cannot identify object attributes, which may prevent mobile robots from taking obstacle avoidance behaviors. Therefore, this article proposes a mobile robot obstacle detection method based on the non-iterative K-means algorithm to identify the attributes of objects in the binocular image and complete the obstacle classification detection. This method uses version 3 of You Only Look Once (YOLOv3) algorithm framework to perform fast target detection in the Microsoft Visual Studio 2005 (VS2015) environment, and uses the non-iterative K-means algorithm to classify the detection target into obstacles and non-obstacles, to achieve rapid detection of obstacles. The results show that the average speed of YOLOv3 is about 30ms, almost 1/3 or even 1/4 times that of other types of detection algorithms (involving SSD (Single Shot MultiBox Detector) for 61ms and R-FCN (Region-based fully convolutional network) for 85ms). The non-iterative K-means algorithm can be used to realize the rapid detection of obstacles, while the RCNN (Region-based Convolution Neural Networks), Fast-RCNN, and SSD have wrong identification.