Ultra Fast Lane Detection

Our new model ZOO works with DepthAI V3. Find out more in our documentation.

0 Likes

Model Details

Model Description

Ultra Fast Lane Detection is a lightweight, fast, and accurate lane detection model. It can detect up to 4 lanes. It can be used in Advanced-driving-assistance-sysyems applications where lane detection is a crucial part. The model must be fast because it is the fundamental component of autonomous driving. The model is based on anchors and the detections are represented as clusters of points.

Developed by: Zequn Qin et.al.
Shared by:
Model type: Computer vision
License:
Resources for more information:

Training Details

Training Data

The model was trained on two lane detection datasets: and . For more information about training data check the .

Testing Details

Metrics

The evaluation was done on the validation set of both training datasets.

Metric	Value
Accuracy	95.87
F1-score	68.4

Results are taken from .

Technical Specifications

Input/Output Details

Input:
- Name: input
  - Info: NCHW BGR un-normalized image
Output:
- Name: output
  - Info: intermediate results (postprocessing is needed)

Model Architecture

Backbone: ResNet-18
Neck: The neck consists of additional layers for feature aggregation and global context learning. The neck takes the feature maps from the backbone and refines them for the lane detection task.
Head: The head is responsible for performing the final task-specific predictions. In this model, the head involves a row-based selecting mechanism for lane detection.

Throughput

Model variant: ultra-fast-lane-detection:culane-800x288

• Input shape: [1, 3, 288, 800] • Output shape: [1, 201, 18, 4]

• Params (M): 44.517 • GFLOPs: 8.396

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	10.52	N/A
RVC4	INT8	379.44	4.09

Model variant: ultra-fast-lane-detection:tusimple-800x288

• Input shape: [1, 3, 288, 800] • Output shape: [1, 101, 56, 4]

• Params (M): 61.221 • GFLOPs: 8.412

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	10.33	N/A
RVC4	INT8	317.18	3.51

* Benchmarked with , using 2 threads (and the DSP runtime in balanced mode for RVC4).

* Parameters and FLOPs are obtained from the package.

Quantization

The RVC4 version of the model was quantized using the Driving dataset in HubAI.

Utilization

Models converted for RVC Platforms can be used for inference on OAK devices. DepthAI pipelines are used to define the information flow linking the device, inference model, and the output parser (as defined in model head(s)). Below, we present the most crucial utilization steps for the particular model. Please consult the docs for more information.

Install DAIv3 and depthai-nodes libraries:

pip install depthai
pip install depthai-nodes

Define model:

model_description = dai.NNModelDescription(
    "luxonis/ultra-fast-lane-detection:culane-800x288"
)

nn = pipeline.create(ParsingNeuralNetwork).build(
    <CameraNode>, model_description
)

Inspect model head(s):

LaneDetectionParser that outputs message (at most 4 clusters of points, each cluster representing one lane).

Get parsed output(s):

while pipeline.isRuning():
    parser_output: Clusters = parser_output_queue.get()

Example

You can quickly run the model using our script. It automatically downloads the model, creates a DepthAI pipeline, runs the inference, and displays the results using our DepthAI visualizer tool. To try it out, run:

python3 main.py \
    --model luxonis/ultra-fast-lane-detection:culane-800x288

Fast and lightweight lane detection model.
License	MIT Commercial use
Downloads	461
Tasks	Line Detection
Model Types	ONNX

Model Variants

Name	Version	Available For	Created At	Deploy
		RVC2, RVC4	9 months ago
		RVC2, RVC4	9 months ago