Research on Computer Vision Models for Deep Learning in Autonomous Mobile Robots

Abstract We propose a computer vision model for semantic and instance segmentation, based on YOLOv8 architecture and partial convolution modules. Transfer learning on the Cityscapes dataset is performed and results show comparable to the SOTA model accuracy. Furthermore, a speed comparison in terms of inference time is conducted, the results of which show that the proposed model is faster when processing larger input images using smaller batch sizes. Our research enhances the potentialities for further improvement in application of fast deep learning models in mobile robotics for complex vision tasks assisting their intelligent behaviour. It also suggests opportunities to increase the suitability of the improved models when implemented on low-resource hardware, such as edge devices and single-board computers, which in turn can facilitate integration of intelligent features and into various autonomous mobile robots in tasks such as self-driving, avoiding and detouring obstacles, optimal trajectory/path selection, increasing safety and reliability.