This paper outlines how we modified the torch2trt library which allowed us to build a recursive framework that can quantize previously unsupported PyTorch models. The framework partitions the PyTorch model into supported and unsupported modules, and then rebuilds the PyTorch model by replacing the supported PyTorch modules with faster TensorRT modules. The framework allows us to optimize and deploy more advanced Deep Neural Network algorithms that are not natively supported by torch2trt.

The research in this paper proposes a novel method for online camera-to-lidar calibration, which aligns the parameters of both lidar and camera sensors during run-time without requiring manual annotations or specialized equipment. The proposed method leverages the robot’s motion and observability constraints to optimize the parameters, resulting in improved accuracy and robustness compared to traditional methods. The approach consists of two main components: a motion-based module that estimates the parameters using the robot’s motion, and an observation-based module that refines the estimated parameters using observability constraints. The method is implemented on popular edge computing platforms and evaluated through experiments in various environments. The results demonstrate the superiority of the proposed method over baseline approaches, showcasing its potential for practical applications. The design of our algorithm and results of its testing on edge computing platforms will be discussed in this paper.