NVIDIA announced Drive Thor on 2022‑09‑22 13:18 UTC as a next‑generation supercomputer for safe and secure autonomous vehicles, supporting assisted driving, parking, driver & occupant monitoring, digital instrument cluster, in‑vehicle infotainment and rear‑seat entertainment.
ZEEKR 极氪 will integrate Drive Thor into its centralized vehicle computer for next‑generation intelligent EVs, with production slated to start in early 2025.
I added a brief overview of visual odometry (VO) as a core vision‑based technique for estimating a self‑driving car’s pose from RGB camera streams. This expands the existing vision‑sensors narrative with the specific goal, inputs, and outputs of VO.
- Goal: Estimate the real‑time trajectory of a self‑driving car.
- Typical pipeline steps:
I chose camera‑only VO for scenarios where GPS is unreliable (e.g., tunnels, urban canyons) and preferred feature‑based methods for computational efficiency on embedded automotive hardware.
- Image acquisition – synchronized camera frames.
- Feature detection & description – detect salient points (e.g., corners, edges).
- Feature matching / tracking – associate features across consecutive frames.
- Motion estimation – solve for relative pose using geometric models.
- Scale recovery – combine with additional sensors or known scene constraints.
- Pose integration – accumulate relative poses to build a trajectory.
- Input: Real‑time RGB images from vehicle‑mounted vision cameras.
- Output: Instantaneous pose (position + orientation) of the vehicle in 3‑D space.
On 2022-09-07 I clarified that Vision‑based ADAS is a cross‑disciplinary area that spans computer vision, machine learning, robotics navigation, embedded systems, automotive electronics, and safety‑critical software.
This framing guides how I compose project teams and prioritize research, ensuring that expertise from each of these fields is integrated into a coherent ADAS architecture.
On 2022-09-09 I promoted a short guide on basic taxonomy for vision‑based ADAS, authored by Jonathan Horgan, Ciarán Eising, John Mc Donald, Senthil Yogamani and other Valeo colleagues, to help new CV engineers get up to speed.
- The guide emphasizes three high‑level autonomous‑vehicle objectives:
- Road safety
- Traffic efficiency
- Driving experience
I decided to use the linked video as the foundational example for illustrating LIDAR vs RADAR capabilities in self‑driving cars.
I examined Tesla’s massive distributed vision training system, Hydranet, which runs on PyTorch and consumes enormous compute resources. The architecture consists of roughly 50 distinct networks, 70 000 GPU‑hours of training, and produces about 1 000 output tensors per inference pass across eight cameras. Optimization highlights - Reuse of backbone models across many CV tasks to cut redundant computation. ^backbone1 - Group related tasks into shared branches, reducing per‑task overhead. ^group1 - Use an unrolled RNN to generate a Bird’s‑Eye‑View (BEV) map that supports curb detection, smart‑summon, and distance estimation. ^bev1 - Selectively omit rear/side camera streams for lane‑line detection to lower bandwidth during inference. ^lane1 - Deploy a “shadow mode” closed‑loop pipeline that feeds live fleet predictions back into training after validation. ^shadowmode1