Lane-Detection-Hybrid-CNN-RNN

This project is part of my undergraduate thesis. I build my own dataset for Indonesian road in three weather conditions: sunny, rainy, night and also three road conditions: straight, turn right, and turn left. The model that I used is based on Zou et. al. [1] which is the combination of UNet/SegNet and LSTM. The lane detection was done in the manner of semantic segmentation. I evaluate the performance of the model using F1-score and compare it with the algorithm that used by my senior in previous undergraduate thesis.

Architecture of the model:

Code and Resources Used:

Thanks to original source code. For the complete dataset for training and original code, you may find it in this link Below is resources for my own implementation

• Python == 3.7.11
• PyTorch == 1.10.2
• Packages: numpy, PIL, sklearn, tqdm, OpenCV

Below is the notebook i used to train the model and process the input video

• Model training :
  • UNet-ConvLSTM
  • SegNet-ConvLSTM
• Video Processing :
  • OpenCV Video Processing

References

[1] Zou, Q., Jiang, H., Dai, Q., Yue, Y., Chen, L., & Wang, Q. (2019). Robust lane detection from continuous driving scenes using deep neural networks. IEEE transactions on vehicular technology, 69(1), 41-54.

[2] Widianto, S. C. (2020). Deteksi Lajur Mobil Otonom Pada Kondisi Gambar Terdistorsi dan Kurang Pencahayaan Menggunakan Pengolahan Citra [Undergraduate Thesis]. Institut Teknologi Sepuluh Nopember.