This paper proposed an efficient road image segmentation algorithm based on multi-scale features. Unlike the existing methods, the introduced model contains two feature extraction branches, namely the Detail Branch and the Semantic Branch. The Detail Branch extracts low-level detail information from the image, while the Semantic Branch extracts high-level semantic information. Then, these extracted multi-scale features from each branch are also fused, which can further improve the model's performance. In addition, we designed a simple and efficient feature fusion module to enhance the model's ability to encode contextual information by fusing multi-scale features. We collected and labeled a large amount of road segmentation data suitable for blind-guiding scenarios and made a corresponding data set. We conduct extensive experiments on the proposed dataset. The experimental results show that the mIoU of the proposed method is 96.5%, which is better than the existing complex models. The structure of our proposed method is shown as follow.

We conduct extensive experiments on our collected dataset. The model performance was evaluated using mIoU. Results are shown as follow:

• The code is developed using python 3.10 on Ubuntu 20.04.
• The proposed method should be trained and validated in GPU.
• If you want to train the proposed segmentation method on custom dataset, Nvidia GPUs are needed.
• This code is development and tested using one Nvidia A100 GPU with 80GB memory.

1. Clone this repo:

   git clone https://github.com/syysha0k/VisNavi
   

2. Install dependencies:

   pip install -r requirements.txt
   

• Training:

  python main.py --model emseg_tiny --vis --device_id 0 --epoches 300 --crop_val --lr 0.01 --crop_size 1024 --batch_size 64 
  

• Test:

  python main.py --model emseg_tiny --vis --device_id 0 --crop_val --crop_size 1024 --batch_size 16 --weight checkpoint/best.pt --test_only --save_path exp/10/test_01
  

• Predict:

  python predict.py --input datasets/test_01.png --model emseg_tiny --weight checkpoint/best.pt --save_path exp/10/pred_01