A Low-Latency Edge AI Framework for Real-Time Road Accident Detection Using Multi-Sensor Fusion and Embedded IoT Systems
DOI:
https://doi.org/10.7492/3drz0d94Keywords:
Edge AI, road accident detection, multi-sensor fusion, IoT, embedded systems, deep learning, real-time inference, MobileNetV3, TinyYOLOv4, MQTT, Kalman filter, low-latencyAbstract
Road traffic accidents remain a critical global safety concern, demanding intelligent, real-time detection systems that can operate reliably under constrained
computational environments. This paper presents a low-latency edge AI framework for real-time road accident detection leveraging multi-sensor fusion and embedded
IoT systems. The proposed architecture integrates data from heterogeneous sensors including accelerometers, gyroscopes, GPS modules, cameras, and LiDAR fused through a
lightweight Kalman-filter-based algorithm optimized for resource-constrained microcontrollers such as NVIDIA Jetson Nano and Raspberry Pi 5. A compressed deep learning model employing
MobileNetV3 and TinyYOLOv4, trained on augmented accident scenario datasets, performs onboard inference with sub-100 ms latency, eliminating cloud dependency for critical decisions.
The framework incorporates an adaptive edge-cloud offloading strategy, ensuring bandwidth efficiency and fault tolerance. Upon accident detection, the system triggers automated emergency
alerts via MQTT-based IoT communication protocols to nearby emergency responders and traffic management centers. Experimental evaluations demonstrate a detection accuracy of 94.7%,
a false positive rate below 3.2%, and an end-to-end latency of 87 ms on embedded hardware. Comparative analysis confirms superiority over existing cloud-dependent and single-sensor
approaches. The framework's scalability and low-power design make it suitable for smart city deployments, autonomous vehicles, and intelligent transportation systems, offering a practical
and deployable solution for proactive accident management.
