ROBOTICS & AVs
Pedestrian & Obstacle Detection
Pedestrian & Obstacle Detection. What it takes to run it.
Safety-critical object classification in crowded environments.
// 01 · The question
“Is there a person in the path, and how far away?”
// 02 · Why it is hard
This is the safety-critical class, so the acceptable miss rate is effectively zero, and it has to hold for people who are partly hidden, seated, crouching or moving fast at the edge of the frame.
// 03 · What the labels have to be
3D cuboid
Decided before the first frame
The stopping decision needs distance, which an image box does not contain. A cuboid gives range and footprint in the robot's frame, and range is what the safety envelope is computed from.
This is the decision that is expensive to reverse. The geometry has to match what the answer contains, and finding out it does not means labelling the set a second time.
// 04 · How we run it
Three parts of one loop, on this job.
LexAnnotate
Pixel-Perfect Data
Pixel-level semantic segmentation, LiDAR point cloud annotation, sensor fusion labeling, and temporal event tracking for complex perception tasks. Expert annotators, accountable for every label.
LexInsight
Spatial Intelligence
Build spatial understanding from flat labels. Map object relationships, track movement patterns, identify zero-day anomalies that standard training data misses.
LexAlert
Perception Alerts
Continuously validate perception performance in production. Detect when sensor drift or environmental changes degrade navigation. Route failures back for retraining before they cause incidents.
// 05 · What breaks it after launch
For outdoor platforms, rain, glare and low sun degrade exactly the sensor the safety case depends on. The conditions that make detection hardest are also the ones that make stopping distances longest.
It is not the only one that can get this use case, it is the one that usually gets it first. All ten conditions.
// 06 · In robotics
Dynamic Spatial Navigation
The Challenge
Flat bounding boxes failed to provide spatial context for autonomous material handling in shifting environments.
Our Execution
Deployed pixel-level semantic segmentation and temporal event tracking for high-fidelity spatial datasets.
The Result
Enabled autonomous systems to adapt to changing environments safely. Reduced manual overrides.
95%+
Navigation reliability in changing environments