Embedded Multi-Layer Safety Framework for Industrial Vehicle Operations Using Deep Learning and Real-Time Monitoring

Escalating deployment of low-speed autonomous vehicles across manufacturing and logistics environments has created urgent demand for safety systems capable of distinguishing human workers from inanimate obstacles in real time. This paper presents a revised and empirically strengthened embedded multi-layer safety framework integrating three-axis ultrasonic obstacle avoidance, transfer-learned SSD MobileNetV2 human detection, YOLOv8n centroid-based activity classification, cooldown-gated text-to-speech alerting, and a Flask supervisory dashboard within a single platform costing under INR 5,000. Extended evaluation introduces mAP@0.5, precision-recall metrics, and memory profiling for both models; an ablation study quantifies the contribution of each layer; and a quantization benchmark demonstrates inference acceleration via TensorFlow Lite INT8 and ONNX INT8 conversion. The SSD MobileNetV2 model achieves 72.4% classification accuracy, mAP@0.5 of 0.58, and a mean IoU of 0.61; YOLOv8n attains 94.0% activity classification accuracy at 24 FPS with mAP@0.5 of 0.89. Full three-layer operation achieves a 96.7% collision avoidance rate and reduces false alert frequency to 1.9 per hour with the cooldown gate active. A 60-minute continuous deployment confirmed sensor precision within 3% error and stable concurrent operation. Limitations regarding industrial dataset coverage, regulatory compliance, and edge-only deployment are discussed alongside a roadmap for future work.