Validate Autonomous Driving Localization Continuity During Sensor Dropout

Overview of Technical Issues:

During sensor dropout events (GPS signal loss in tunnels, LiDAR degradation in adverse weather), the sensor fusion module insufficiently integrates remaining data sources and the localization computing module inadequately maintains positioning estimates, resulting in interrupted or unreliable positioning output transmitted to the vehicle control system as a harmful effect that threatens autonomous navigation safety; the goal is to ensure continuous localization accuracy within acceptable thresholds throughout all sensor dropout scenarios.

Solution directions generated for this problem

Problem Direction 1 :

ImproveSensor fusion integration capability
VS
ConstraintComputational processing load

Inspiration 1 : Cross-domain reference

Application Principle: #35 Parameter changes
Cross-domain applicability Assess applicability
Generation of interactive audio tracks from visual content
Innovative Solution Refine solution

Event-triggered adaptive fusion mode switching for sensor dropout resilience

Switch fusion modes based on sensor health
How to solve :
  • Implement three-tier fusion architecture: Tier-1 (GPS+IMU, 38-42% CPU) for normal operation, Tier-2 (GPS+IMU+odometry+map-matching, 58-65% CPU) for single sensor dropout, Tier-3 (IMU+odometry+visual odometry+vehicle dynamics, 72-78% CPU) for multi-sensor dropout
  • Deploy event-driven mode controller monitoring sensor health flags at 10Hz — trigger Tier-2 when any primary sensor confidence drops below 70%, trigger Tier-3 when two or more sensors fail, revert to lower tier within 2 seconds after sensor recovery
  • Pre-load fusion algorithm parameters in lookup tables indexed by sensor availability patterns — 12 common dropout scenarios (tunnel entry, heavy rain, urban canyon) with pre-computed Kalman gain matrices and covariance bounds, reducing runtime calculation by 85% during transitions
Expected Effect : Average CPU 45-52% vs constant 75-85%; positioning continuity 98.5%; error <1% for 35+ seconds during dropout
Risk Control :
  • mode transition latency exceeding 500ms causing estimate jumps
  • lookup table coverage insufficient for rare scenarios
  • sensor health flag false positives triggering unnecessary mode switches

Problem Direction 2 :

ImproveLocalization estimate stability
VS
ConstraintSensor measurement precision requirement

Inspiration 1 : Cross-domain reference

Application Principle: #26 Copying
Cross-domain applicability Assess applicability
Tire wear prediction system, tire wear prediction program, tire wear prediction method, and data structure
Innovative Solution Refine solution

Virtual sensor model fusion for drift-free localization

Fuse physical sensors with virtual models
How to solve :
  • Deploy vehicle dynamics virtual sensor generating synthetic odometry from steering angle, throttle, and yaw rate inputs using bicycle kinematic model (update rate 50Hz, latency <10ms)
  • Implement tire-road interaction model calculating slip ratio and longitudinal force to correct wheel encoder drift, achieving effective 0.8% odometry accuracy from 3% hardware through physics-based compensation
  • Establish statistical fusion framework weighting physical IMU (10°/hr bias) at 60% and virtual dynamics model at 40% during GPS dropout, maintaining positioning error <1.2% over 30+ seconds without hardware upgrade
Expected Effect : Error <1.2% for 30s; no sensor upgrade; cost neutral
Risk Control :
  • model parameter calibration drift
  • tire wear affecting virtual accuracy
  • computational validation complexity

Problem Direction 3 :

ImproveSensor fusion integration capability
VS
ConstraintSystem implementation complexity

Inspiration 1 : Cross-domain reference

Application Principle: #1 Segmentation
Cross-domain applicability Assess applicability
Techniques and apparatuses for multiplexing schemes for millimeter wave downlink single carrier waveforms
Innovative Solution Refine solution

Modular sensor-specific estimator architecture with standardized fusion hub

Partition fusion into independent modules
How to solve :
  • Decompose fusion system into independent sensor-specific estimators (IMU estimator, odometry estimator, map-matching estimator, visual odometry estimator) that operate autonomously without cross-dependencies
  • Each estimator publishes standardized pose messages (position, velocity, uncertainty covariance, timestamp) at defined rates (IMU 100Hz, odometry 50Hz, map-matching 10Hz) to a central fusion hub via publish-subscribe protocol
  • Fusion hub implements federated Kalman filter architecture where each sensor maintains local state estimate
  • hub combines estimates using information-weighted averaging without requiring inter-estimator communication, reducing coupling by 65%
Expected Effect : Add visual odometry with 1 new module only; validation scenarios reduced from N² to N; integration time <2 weeks per new sensor
Risk Control :
  • estimator output synchronization drift
  • fusion hub message queue overflow during peak load
  • standardized interface insufficient for exotic sensor types

Problem Direction 4 :

ImproveLocalization estimate stability
VS
ConstraintComputational processing load

Inspiration 1 : Cross-domain reference

Application Principle: #10 Preliminary action
Cross-domain applicability Assess applicability
Systems and methods for identifying cancer treatments from normalized biomarker scores
Innovative Solution Refine solution

Pre-computed Trajectory Library for Dropout-Resilient Localization

Build trajectory library offline for dropout resilience
How to solve :
  • During GPS-healthy periods (CPU at 40%), pre-compute and cache 30-second prediction trajectories for 12 standard maneuver types (straight, left/right turns at 15°/30°/45°, lane changes, accelerations at ±2m/s²) using full extended Kalman filter with IMU-odometry fusion, storing position, velocity, and error covariance matrices at 0.5s intervals
  • Upon dropout detection, match current vehicle state (speed ±5km/h, steering angle ±3°, acceleration ±0.5m/s²) to nearest cached trajectory template via k-nearest neighbor lookup (k=3, execution time <2ms), then apply lightweight linear interpolation and scaling to adapt pre-computed predictions to actual motion, reducing real-time CPU load from 75-85% to 48-52%
  • Implement rolling cache update mechanism that refreshes 10% of trajectory library every 5 seconds during normal operation, ensuring templates reflect current IMU bias estimates (updated via recursive least squares) and road conditions, maintaining prediction accuracy within 0.8% distance-traveled error for 30+ seconds
Expected Effect : Error <1% for 30s, CPU 48-52% vs 75-85%, cache size 15MB
Risk Control :
  • trajectory library coverage gaps for rare maneuvers
  • cache staleness if IMU bias drifts rapidly
  • lookup latency spike under memory contention

Problem Direction 5 :

ImprovePositioning output continuity
VS
ConstraintComputational processing load

Inspiration 1 : Cross-domain reference

Application Principle: #11 Beforehand cushioning
Cross-domain applicability Assess applicability
Window-based scheduling using a key-value data store
Innovative Solution Refine solution

Pre-computed fallback trajectory library for zero-overhead dropout recovery

Pre-compute dropout recovery trajectories during idle cycles
How to solve :
  • During GPS-healthy periods (CPU at 40%), execute background process to pre-compute 30-second fallback trajectory libraries covering 12 standard maneuvers (straight, left/right turns at 15°/30°/45°, lane changes, acceleration/deceleration profiles) using full extended Kalman filter with IMU+odometry fusion, storing results in indexed lookup tables (≤50MB memory)
  • Upon dropout detection, pattern-matching engine identifies current vehicle state (velocity 0-120 km/h in 10 km/h bins, yaw rate ±30°/s in 5°/s bins) and retrieves nearest pre-computed trajectory within 2ms, adapting it via lightweight linear scaling (3% CPU) instead of real-time computation
  • Implement dual-buffer architecture—primary buffer serves active lookups while secondary buffer updates every 5 seconds during normal operation, ensuring library freshness without interrupting retrieval
Expected Effect : CPU during dropout 45-50% vs 75-85%; output continuity 99.2%; positioning error <1% for 30s; retrieval latency <5ms
Risk Control :
  • maneuver library coverage gaps for rare scenarios
  • pattern-matching misidentification under abrupt maneuvers
  • memory overflow if library expands beyond allocation

Problem Direction 6 :

ImprovePositioning output continuity
VS
ConstraintSystem implementation complexity

Inspiration 1 : Cross-domain reference

Application Principle: #11 Beforehand cushioning
Cross-domain applicability Assess applicability
Redundant braking system having pressure supply for electric vehicles and vehicles having autonomous driving of level 3 (HAD) to level 4 (FAD)
Innovative Solution Refine solution

Pre-computed fallback mode architecture with instant-switch positioning continuity

Pre-compute positioning fallback modes offline
How to solve :
  • During system initialization, pre-load three predefined fallback algorithms (GPS+IMU fusion, IMU+odometry dead-reckoning, pure odometry extrapolation) into reserved memory partitions with pre-calculated Kalman filter matrices and error covariance tables for typical driving scenarios (straight, turn, acceleration)
  • Implement a lightweight health scoring module (5% CPU) that monitors sensor signal-to-noise ratio and data freshness every 100ms, triggering instant algorithm pointer switch when GPS drops below −140dBm or LiDAR point density falls below 200 points/frame, with transition latency under 20ms
  • Each fallback mode operates independently with pre-validated accuracy envelopes (Mode 1: ±0.15m, Mode 2: ±0.5m for 30s, Mode 3: ±2m for 10s) stored as lookup tables indexed by vehicle speed and maneuver type, eliminating runtime state machine complexity and reducing validation test cases from 2^N combinations to 3 fixed scenarios
Expected Effect : Output continuity 99.2%, CPU +8%, validation scenarios reduced 75%
Risk Control :
  • pre-computed matrix accuracy degradation over software updates
  • memory partition overflow during edge-case maneuvers
  • transition latency spike under concurrent sensor failures
Patsnap Eureka Solution