Semantic Segmentation Approach on The Off-Road Vehicle Navigation Dataset Using Camera Vision

In this study, a deep learning network for semantic segmentation based on the Deeplabv3+ network is applied to the set of images to form a dataset. A total of 1400 images were captured from a bird's eye chili farm using a stereo camera. The dataset is separated into a 60:20:20 ratio for training, testing, and validation. The network performance is determined by the result of the segmented image's accuracy and Intersection-over-Union (IoU) tested on the testing images. Based on the visual observation of the outcome, it is concluded that the Deeplabv3+ network can identify the pathway and obstacles in the image of the chili farm environment. The training accuracy obtained is 93.18%. Global accuracy achieved is 0.92781, mean accuracy is 0.88412, mean IoU is 0.75869, weighted IoU is 0.87017, and mean boundary F1 (BF) score is 0.79606.