Multi-Modality Sensing Powered by Laser Dynamics: Theory and Applications

The continued development of embodied intelligence and autonomous driving technologies necessitates comprehensive multi-dimensional environmental perception. Conventional discrete sensor fusion paradigms, such as the physical integration of LiDAR, visual cameras, and displacement transducers, face constraints regarding Size, Weight, Power, and Cost (SWaP-C), while also introducing engineering challenges in dynamic environments, including parallax errors, spatiotemporal misalignment, and long-term calibration drift. To address these limitations, this thesis proposes a unified perception architecture utilizing native hardware multiplexing. By combining the nonlinear dynamics of a semiconductor laser with computational imaging algorithms, this paradigm establishes a framework for executing multi-dimensional perception through a single physical aperture.At the physical hardware layer, the dynamical evolution mechanisms of the semiconductor laser under varying external optical feedback conditions are systematically investigated to realize high-fidelity one-dimensional (1D) sensing modalities. Initially, an intrinsic linear displacement sensing modality is constructed within the steady-state, strong optical feedback regime. This configuration suppresses the mode-hopping hysteresis characteristic of traditional self-mixing interferometry and mitigates the computational latency associated with phase unwrapping algorithms, providing a basis for real-time tracking of high-frequency mechanical vibrations. Progressing beyond the steady state, the system is driven past the Hopf bifurcation boundary into the Period-One (P1) nonlinear oscillation regime, transforming the laser into a microwave photonic sensing modality. To mitigate parameter optimization difficulties and chaotic sensitivity in this state, a machine-learning-assisted control framework is implemented. The application of a Feedforward Neural Network (FNN) surrogate model allows for deterministic parameter control, generating a stable microwave carrier suitable for Frequency Modulated Continuous Wave (FMCW) ranging.To expand the system’s capability from 1D point detection to two-dimensional (2D) spatial perception, a dual-track computational imaging framework is formulated to retrieve spatial target information from one-dimensional intensity fluctuations. Standard Approximate Message Passing (AMP) algorithms typically diverge when applied to the highly correlated, non-independent and identically distributed physical diffraction matrices inherent in optical systems. To resolve this limitation, a Phase-Retrieval Unitary Approximate Message Passing (PR-UAMP) algorithm is utilized. By preconditioning the measurement matrix through singular value decomposition, the algorithm restores the necessary statistical stability for iterative reconstruction. Within this framework, Sparse Bayesian Learning (SBL) is applied for inherently sparse targets, and orthogonal Haar wavelets are selected for dense natural images to maintain unitary invariance. Both numerical simulations and optical experiments verify that this framework achieves accurate phase retrieval and computational stability under sub-Nyquist sampling and measurement noise conditions.At the system integration layer, these physical sensing and computational imaging modalities are synthesized into a dual-task hardware multiplexing architecture suitable for autonomous vehicles. By incorporating a high-speed optical switch and a Time-Division Multiplexing (TDM) strategy, the independent modalities are unified into a shared coaxial optical path. This integrated architecture physically aligns the multi-modal data acquisition, inherently reducing spatiotemporal synchronization errors and optimizing the SWaP-C metrics of the perception system. Ultimately, this research provides a consolidated physical mechanism and algorithmic foundation for the development of highly integrated, multi-dimensional cognitive perception systems.