Real-time Semantic Segmentation on Resource-Constrained Robotic Applications

This thesis tackles the challenge of implementing semantic segmentation, a data-intensive and computationally demanding process, in resource-constrained scenarios such as edge devices. The study centers on a dataset collected by Stiga from a lawn mower robot operating in a garden environment. The unique challenge for algorithms lies in the varying outdoor lighting and weather conditions, particularly due to the scarcity of data related to extreme weather conditions, which leads to the model’s performance degradation. The primary objective of this thesis is to achieve a target of 10 fps while maintaining satisfactory performance. The architectures of U-Net and DeepLabV3 were adapted, with the computationally intensive encoder components replaced by efficient blocks from MobileNetV2 and MobileNetV3. These blocks, pre-trained on ImageNet, can generate a hierarchy of high-level features that the decoder uses to produce an accurate segmentation map. The challenge posed by varying weather conditions is mitigated to some extent through data augmentation. To validate the effectiveness of the models designed in this study, five additional datasets, similar to the garden environment, were employed. The results reveal that training the models on both the encoder and decoder components led to enhanced performance. Specifically, the DeepLabV3 model with a MobileNetV2, encoder of depth 4 achieved superior performance, with a MeanIOU of 95.5% and a frame rate of 15.84 fps on the Stiga dataset. Furthermore, these models also demonstrated superior performance when applied to other datasets similar to the Stiga dataset.