AutoSpeed - In-path objects detection
For autonomous cruise control applications, it is crucial to maintain a safe following distance from the vehicle in front, known as the closest-in-path object. Detecting the closest-in-path object therefore becomes a very important task for any self-driving or driver assitance application, as it also supports important safety features such as forward collision warning and autonomous emergency braking. The AutoSpeed network is a bounding box detection model inspired by the YOLOv11 architecture, in which the backbone c3k2 blocks are substitued by a custom designed 'context' block for improved overall scene understanding.
The AutoSpeed model detects all foreground objects and classifies objects into three categories depending on the object's position with respect to the predicted future driving path of the ego-car:
- objects directly within the future driving path of the ego-car
- objects cutting-in/cutting-out of the future driving path of the ego-car
- objects outside of the future driving path of the ego-car
Performance Results
The AutoSpeed 2.0 network is trained on the OpenLane dataset on the CIPO detection task. We modify the ground-truth labels by merging 'Level 3' and 'Level 4' category objects in the ground truth dataset into a single class. We use a 90:10 ratio for train:val split. The evaluation metrics are summarized in the table below:
| mAP@50 | mAP | |
|---|---|---|
| FP32 | 0.74 | 0.56 |
| INT8 | 0.73 | 0.54 |
Model variants
The AutoSpeed network is trained in two variants, the original AutoSpeed 1.0 network processes frames in square aspect ratio with size 640px by 640px. AutoSpeed 2.0 processes frames in a 2:1 aspect ratio with size 1024 px by 512px, allowing for objects to be detected further away and a wider viewing angle of the scene which is more suited to autonomous driving applications. We recommend you work with AutoSpeed 2.0 for new applications and developments.
| Input frame | PyTorch | ONNX FP32 | ONNX INT8 | |
|---|---|---|---|---|
| AutoSpeed 2.0 | 1024x512 (2:1) | model weights | model weights | model weights |
| AutoSpeed 1.0 | 640x640 (1:1) | model weights | model weights |
Quick start
from PIL import Image
from huggingface_hub import hf_hub_download
from Models.inference.auto_speed_infer import AutoSpeedNetworkInfer
repo_id = "AutowareFoundation/auto_speed"
ckpt = hf_hub_download(repo_id, "models/autospeed_2.pth")
model = AutoSpeedNetworkInfer(ckpt) # uses CUDA if available
TEST_IMG = hf_hub_download(repo_id, "assets/highway.jpg")
img = Image.open(TEST_IMG).convert("RGB")
detections = model.inference(img)
print(detections)
Citation
@software{autoware_auto_speed_2026,
author = {The Autoware Foundation},
title = {AutoSpeed: In-path object detection model},
year = {2026},
version = {2.0},
url = {https://github.com/autowarefoundation/auto_speed},
note = {Checkpoints: https://huggingface.co/AutowareFoundation/auto_speed}
}