Embedded intelligent vehicle perception and control using neural network methodologies
TL;DR
The proposed platform provides an explainable and cost-effective engineering framework for embedded intelligent vehicles in educational and inspection scenarios, balancing algorithmic efficiency with practical deployability.
Abstract
This paper presents the design and implementation of an embedded intelligent vehicle system for resource-constrained environments. The system addresses key challenges in environmental perception, multimodal control, and dynamic obstacle avoidance. Built on the OpenMV Cam H7 platform, it deploys a quantized MobileNetV2 model optimized with TensorFlow Lite for real-time traffic-sign recognition while maintaining computational efficiency. The overall architecture integrates an STM32 microcontroller with a proportional-integral-derivative-based motion-control strategy to support multi-stage braking. To satisfy diverse operational requirements, the vehicle also provides Bluetooth remote control and voice-command interfaces, supported by a priority-based arbitration mechanism that enables safe transitions between control modes. Environmental navigation is enhanced through visual detection and adaptive path-marking recognition, enabling dynamic target tracking and accurate ground-path perception. For collision avoidance, the system fuses OpenMV visual data with ultrasonic and infrared sensing to implement a hierarchical avoidance strategy, including proportional-integral-derivative-controlled emergency braking in critical situations. Experimental results in simulated road environments show robust real-time performance. The proposed platform provides an explainable and cost-effective engineering framework for embedded intelligent vehicles in educational and inspection scenarios, balancing algorithmic efficiency with practical deployability.