Low power indirect time of flight near field LiDAR depth correction

Indirect time-of-flight (iToF) LiDAR sensors have their technical challenges. They are a viable alternative to direct time-of-flight (dToF) LiDAR sensors and cameras for near-field applications, particularly in autonomous vehicles and mobile robots. iToF LiDAR sensors face issues with depth range ambiguity. Many modern Deep Neural Network (DNN) computer vision models that run on graphics processing units can effectively address this problem. However, GPUs' high electrical energy consumption poses a significant challenge when integrated with battery-based embedded systems. Nonetheless, viable alternatives can provide sufficient computational power while maintaining a low electric power profile. One such option is the low-power ARM processor architecture, which is supported by single-instruction, multiple-data (SIMD) accelerators. In this paper, we go deep into the system architecture differences between GPUs and CPUs-integrated SIMD, exploring the potential use cases for running DNN models, identify which models are suitable for which architectures, propose methods for running the depth correction algorithm on such low-power platforms, discuss the challenges of these methods, and suggest ways to overcome them, thereby enabling the use of sensors in low-power applications. We pose several questions about the embedded deployment of DNN models and offer recommendations to address them. We were able to achieve a real-time depth correction rate at 10 frames per second at acceptable accuracy, with no need for GPUs on low-power ARM architecture Cortex-A76.