Abstract
Traffic sign recognition is a safety-critical perception task in advanced driver assistance systems and intelligent transportation platforms. Although convolutional neural networks (CNNs) provide strong classification performance, their deployment on low-cost embedded devices is constrained by limited memory, compute capacity and inference latency. This paper presents an edge-based traffic sign recognition system that combines a lightweight MobileNet-based CNN with post-training 8-bit integer (INT8) quantisation for Raspberry Pi-class deployment. The system is evaluated on the German Traffic Sign Recognition Benchmark (GTSRB), which contains 43 traffic sign classes. The floating-point model achieves 93.57% accuracy, while the INT8 quantised model achieves 93.49% accuracy, corresponding to only 0.08 percentage-point accuracy degradation. The quantised model also reduces arithmetic precision from 32-bit floating point to 8-bit integer representation and achieves an average inference latency of 0.28 ms per image compared with 0.50 ms for the floating-point model. The results show that post-training quantisation preserves recognition performance while improving edge inference suitability. The study demonstrates that quantised CNNs can support real-time, privacy-preserving and cloud-independent traffic sign recognition on resource-constrained embedded hardware.