Towards Accurate Multisensor Registration – A Pattern Optimization Approach

The increasing complexity of autonomous vehicles necessitates the integration of multiple perception sensors for robust and reliable environment understanding. However, the full potential of these sensors is limited by the absence of precise extrinsic calibration between them. This paper presents a novel nested optimization algorithm for calibration target pattern design to achieve accurate multisensor registration. The proposed framework combines: (i) a Genetic Algorithm (GA) for determining the optimal number of calibration targets; (ii) constrained nonlinear optimization (fmincon) for refining target coordinates; (iii) a full-factorial Design of Experiments (DoE) for structured initial placement; and (iv) Horn’s quaternion method for estimating sensor-to-body transformations. Gaussian noise is simulated in accordance with RADAR datasheet specifications to mimic real-world inaccuracies.The method is validated on a simulated setup with two long-range RADARs (LRRs) and two short-range RADARs (SRRs). The optimized target arrangement (17 targets) reduced the overall transformation RMSE from 8.7cm to 0.74cm for the worst-case sensor and achieved sub-centimeter accuracy for the remaining sensors, corresponding to a relative improvement of over 90%. These results demonstrate the practicality and robustness of the proposed approach for autonomous vehicle sensor calibration, with potential extension to LiDAR and camera systems.