Using 2D LiDAR on a Following Cart: Obstacle Detection & Sensor Fusion

Autonomous following on flat industrial floors comes down to a direct clearance problem: knowing where the operator is moving, and keeping enough clear stopping distance along the vehicle’s trajectory under full payload.

A 2D LiDAR measures visible obstacles across one horizontal plane. That slice captures vertical surfaces like legs, carts, and walls cleanly, but it also has physical limits: it cannot see beneath its mounting height, detect floor drops, or look through stacked loads. Measuring, filtering, and physical braking all take time. Here is how raw range readings translate into speed limits and braking zones on a following cart.

One Laser, Sweeping a Single Horizontal Plane

On flat indoor surfaces, common obstacles — human legs, wheels, upright crates, building pillars, and walls — rise vertically from the floor. A single horizontal beam mounted at lower-leg height intercepts these surfaces directly, providing geometric range measurements with minimal data overhead.

The time-of-flight measurement loop is direct:

1. Emit.A laser diode fires a short infrared light pulse.
2. Reflect.The pulse strikes an obstacle surface and returns.
3. Measure.A photodetector catches the reflection, and the sensor clocks the round-trip flight time (Δt).
4. Calculate.Distance = (speed of light × round-trip time) ÷ 2.
5. Rotate.The optics spin continuously, repeating measurements across sequential angular steps.

At 10 Hz, the scanner takes approximately 100 ms to complete a single 360° revolution. This means any specific direction is revisited roughly once every 100 ms. If an obstacle enters the path immediately after the beam sweeps past, it can take up to a full cycle before the first return is recorded. Signal filtering, inter-module communication, and controller logic add further latency before a braking command reaches the motor drivers — after which the loaded cart still requires physical stopping distance.

Turning Raw Scans into Safe Speed Limits

Each sweep arrives as an array of measured ranges at sequential angles. To translate that array into safe motion commands, the controller evaluates three sequential stages:

Filtering involves an inherent latency trade-off. Airborne dust can produce transient returns, while retroreflective materials can cause measurement artifacts. Requiring confirmation across consecutive scans can reduce some false detections, but it also delays detection and may suppress small or briefly visible obstacles. Filter persistence must be tuned against the minimum obstacle size and reaction budget required for the cart’s operating environment.

Clustering provides geometry, not semantic labels. Nearby points are segmented into candidate obstacles. Return density varies with distance, surface angle, and reflectivity: a target close to the scanner produces a dense cluster of points, while the same object farther away returns only a few points. Tracking a cluster across sweeps in the cart frame estimates relative motion; estimating a world-frame velocity vector requires compensating for the vehicle’s own wheel odometry and yaw rate.

Clearance fields must budget for physical braking distance. The stop field must extend beyond the cart footprint by the distance travelled during sensing and control delay, plus the physical braking distance and a clearance margin. These distances must be validated at permitted speed and rated payload, on the lowest-traction surface and steepest slope allowed for operation. As clearance decreases, the controller lowers permitted speed so the cart can still stop within the available distance. Entry into the stop field issues a braking request.

Why 2D LiDAR Fits Walking-Speed Following

For a vehicle trailing a person at walking speeds on flat terrain, a 2D LiDAR offers clear operational strengths:

• Not dependent on ambient illumination. The sensor carries its own active infrared emitter and calculates distance directly from pulsed flight time, operating in unlit corridors and dim warehouses without external lights. Strong sunlight can reduce measurement performance, and highly reflective surfaces can introduce measurement artifacts depending on sensor design.
• No camera imagery. The scan records geometric range and angle rather than photographic images, simplifying privacy considerations in occupied facilities.
• Manageable compute budget. A 2D scan is significantly less demanding to process than dense 3D point clouds. The controller still requires adequate processing margin for filtering, tracking, and motor control, with worst-case response times measured under full system load.

Physical Constraints and Edge Cases

A single 2D laser plane has clear physical constraints that must be handled in system design:

• Single-plane coverage. Obstacles strictly above or below the mounting height — such as empty 144 mm EPAL pallets, fork tines, descending steps, or overhanging tabletops — will not intersect the beam. Auxiliary sensors (such as ultrasonics or bumper switches) can help cover specific vertical zones, but their mounting angles, response times, and blind spots must be verified on the assembled cart.
• Specular surfaces and clear glass. Clear architectural glass allows 905 nm infrared pulses to pass through or deflect away, meaning the LiDAR may only register the wall behind the door. Where glass partitions exist, path planning and vehicle operating rules must account for optical transparency.
• Adverse outdoor weather and terrain. Heavy rain, fog, and airborne dust scatter optical pulses. Dynamic filters help suppress transient scatter, but severe weather degrades return validity, requiring conservative speed caps. If valid measurements no longer cover the required stopping field, the cart must stop. Furthermore, outdoor pitch and roll tilt the scan plane into the ground or upward into open air.

LiDAR + UWB: Target Tracking Meets Obstacle Clearance

Following requires two distinct pieces of information: where the target person is, and what lies in the vehicle’s path. An ultra-wideband (UWB) subsystem tracks the operator’s tag: Two-Way Ranging (TWR) provides range measurements, while a multi-antenna receiver estimates relative bearing from Phase Difference of Arrival (PDoA). Range and bearing have different error profiles, and human body shielding can attenuate or reflect radio signals in non-line-of-sight conditions.

The two sensors serve complementary roles: UWB establishes the target pursuit vector, while the 360° LiDAR measures obstacle geometry in the surrounding plane. Because a 2D LiDAR point cloud carries no inherent identity labels, associating a laser cluster with the target person requires spatial gating and temporal tracking. The controller checks requested motion against clearance fields and requests deceleration or braking when detected obstacles cross configured thresholds.

Operational Scenarios and Design Assumptions

When applying 2D LiDAR to mobile carts, field performance depends on operating conditions and defensive assumptions:

Illustrative scenario: Industrial aisles. Consider a cart following an operator along a warehouse aisle with a nominal 1.2-meter trailing gap. When cross-traffic (such as a crossing pedestrian or hand truck) enters the forward corridor, the controller initiates deceleration or braking according to the budgeted stopping distance. Resuming motion requires valid scans to show the monitored corridor clear over the configured confirmation interval, rather than restarting immediately when a single ray registers open space.

Illustrative scenario: Outdoor pathways. On paved paths and level fairways, the controller follows the leader’s UWB tracking vector while monitoring for static obstacles like benches, trees, and other carts. Here, terrain grade changes, curb drops, and vibration introduce pitch shifts that can tilt the scan plane into the ground or upward into open air. Such environments require strict speed limits, level operating surfaces, and complementary sensors for drop-offs.