Towards autonomous vehicles: object detection with LIDAR U-net model

Self-driving technology is being developed simultaneously by different players of the automotive ecosystem, ranging from big conglomerates to small startups. Advancing the technology needed for autonomous vehicles is perhaps one of the most intriguing challenges for experts in data analysis and computer vision.

Our AI Engineer Joni Juvonen has a vast background in computer vision. He is quite certain he has taken part in most computer vision related competition at the machine learning community Kaggle. For him, it is the best way to learn while getting hands on interesting real-world data. On the other hand, for the companies offering these challenges, an online competition is an excellent way to tap into the internet’s potential of eager machine learning experts that are ready to help solve the problems.

One such challenge is the Uber competitor Lyft’s recent challenge on Object Detection for Autonomous Vehicles. The task was to build and optimize algorithms based on a large-scale dataset. The dataset Lyft open-sourced for the challenge consisted of the raw sensor camera inputs as perceived by a fleet of multiple, high-end, autonomous vehicles in a restricted geographic area.

The results of Lyft’s object detection challenge were finalized earlier this week, awarding a bronze medal for Joni’s LIDAR U-Net model. His model stood at the 53rd position out of 550 competing teams.

“The challenge in this competition was to leverage data from various different data sources: LIDAR, cameras and the map of the area.”, says Joni.

However many competitors like Joni, chose to use just one LIDAR sensor, located on top of the vehicle. LIDAR stands for light detection and ranging and is an optical remote-sensing technique that uses a pulsed laser to sample the surroundings: “Although I only used the vehicle's top LIDAR sensor for object detection, it's amazing what you can "see" with just that”, Joni comments.

“During the competition, I briefly tried PointRCNN and second.pytorch's PointPillar models, and these both seemed very promising. Due to limited time, I went with my initial trained-from-scratch 5-layer U-Net with Mish and Radam and applied some object tracking and stationary vehicle detection as post-processing.”

“LIDAR is quite new and exotic data format in these kinds of competitions. It was exciting to get the opportunity to test it in use.”

In this video, the 3D boxes drawn around vehicles (on the left side) are Joni’s model's predictions based only on the top-view LIDAR projection (on the right side).

Link to Guido Zuidhof's reference model which Joni used as a starting point for his model:

Interested in hearing how Silo.AI and its machine learning experts can improve your way towards automated operations in your business? Get in touch at

Pauliina Alanen
Former Head of Brand
Silo AI
Share on Social
Subscribe to our newsletter

Join the 5000+ subscribers who read the Silo AI monthly newsletter to be among the first to hear about the latest insights, articles, podcast episodes, webinars, and more.

Pauliina Alanen

Former Head of Brand

Silo AI

What to read next

Ready to level up your AI capabilities?

Succeeding in AI requires a commitment to long-term product development. Let’s start today.