Multiscale Feature Extraction Network for Real-time Semantic Segmentation of Road Scenes On the Autonomous Robot

Semantic segmentation is an effective means for autonomous robots to understand the surrounding scenes. For autonomous robot, it requires the balance of accuracy and speed. Moreover, it is necessary to correctly extract environmental information in complex environments such as occlusion, poor illumination, and shadows condition. To solve above problems, a novel image-based Multi-scale Feature Extraction Network (MFENet) is designed for real-time semantic segmentation task. This network preserves different level features in the encoder and fuses those features to accurately segment each object. In addition, to enhance the representation ability, fusion module is introduced for information exchange between feature maps with different spatial resolution. Moreover, standard convolution is replaced by Multiscale Feature Extraction (MFE) module in intermediate layers, which could strengthen the feature extraction ability. On the Cityscapes dataset, MFENet achieves 72.4% Mean Intersection over Union (MIoU) with 8.0 million parameters at the speed of 30.5 FPS on a single GTX 1070Ti card. Finally, MFENet is deployed on an autonomous robot and tested in the real world. It produces good semantic segmentation results at the speed of 65.5 FPS. The results reveals the proposed MFENet could work well in real-world applications.