Asad-Ismail / lane_detection

Lane Detection and Classification using Front camera monocular images

🚗 Ego and Side Lane Detection for ADAS Applications (Tensorflow 2.x) 🚘

🌟 Features

1. Segement and classify ego, left and right lane lines. The output can be visualized as shown below:

2. Apply pruning, clustering, and quantization to miniaturize the model, making it embedded system-ready.

3. C++ inference to use the resulting miniature model.

📚 Training Dataset:

• The training data is TU Simple Lane Detection dataset.You can access it through this link: https://github.com/TuSimple/tusimple-benchmark

• Dataset is preprocessed to annotate the lanes in 4 categories left ego (label=2), right ego (label=1), right lane (label=3), and left lane line (label=4)

• Images are 1280x720 RGB, and labels are 1280x720 grayscale images

• Image augmentations like rotation, flipping, saturation, brightness, and contrast changes are applied randomly

• Data pipeline is made efficient using data interleaving and prefetch

🧰 Model:

• MobileNetV2 is used as the backbone network, and then transposed convolutions are applied for upsampling with UNET-like feature concatenation.

• ResNets or EfficientNets can also be used as a backbone for better performance.

💻 Training and Prediction

• Install requirements using pip install -r requirements.txt

• Run training using python train.py --train_images [path to train images] --train_labels [path to train labels]

• Perform prediction on image or video using pred.py or pred_video.py (also writes the blended video).

• Prune, cluster, and quantize model weights and activations for miniaturization.

  

Applying the above pipeline reduces the model size by approximately 11x, from 24MB to 2.2MB.

• Pretrained weights are available at https://drive.google.com/drive/folders/1EhQ-8UoFv4rvMqe2mrJ4HFzZATd_Ee8c?usp=sharing.