Abstract: The existing real-time semantic segmentation algorithms for embedded terminals have weak processing capabilities for object detailed features. A method of fusing spatial features of different levels is proposed to crack the above nut. Based on the modified ENet, the inverted residual structure is used in the down-sampling layer to increase the image information acquisition in the network calculation process and decrease the loss of spatial image features caused by the down-sampling process. The spatial attention mechanism is used to weight the down-sampled image spatial feature information to enhance relevant features and weaken irrelevant features. This method connects the low-level high-resolution spatial features to the deep layers of the network. It merges them with in-depth semantic features, which improved the image detail processing capability of the network. Experiments on NVIDIA Jetson TX2, NVIDIA Jetson Xavier NX and NVIDIA Jetson Xavier AGX show that the proposed network runs the same speed as ENet. The mean Intersection of Union (mIoU) on the Cityscapes is increased by 2.9%, and the mIoU on the CamVid is improved by 3.2%.