VGG16 and MobileNet Performance Evaluation on Edge Device in Self-Driving Car Technology

摘要

An end-to-end methodology for training convolutional neural networks (CNN) is proposed in this paper for multi class classification of mobile robots using pre-trained weights. Object and pedestrian identification for safe navigation in self driving cars is the main challenge tackled using the proposed strategy. Modern deep learning-based object detectors significantly progressed in detecting and classifying objects from camera vision, However, hardware constraints necessitate a lightweight design. Therefore, two approaches are proposed: First, the use of the MobileNet architecture, which is intentionally designed for lightweight and computing efficiency in comparison to other architectures, and secondly Pre-trained weights from VGG16 learned on the large-scale ImageNet dataset are used to aid the efficient calculation of the proposed MobileNet architec ture and others. A specialized virtual environment is developed to maintain computational integrity and isolate the local environ ment from interruptions. This virtual environment is rigorously tailored to satisfy the precise requirements of the architectures being built and evaluated. The local environment is not disrupted, and all necessary dependencies are installed effortlessly within the virtual environment. The suggested network’s efficacy is proven by experimental tests on a scaled model of Quanser’s latest self driving car, equipped with the Nvidia Jetson TX2 platform.