Abstract
Recent cars have been equipped with many technologies to raise the safety level of vehicles and pedestrians. Advanced driver assistance systems (ADAS) are intelligent systems based on a huge number of sensors and actuators. Due to safety concerns, traffic sign detection is one of the important systems in ADAS. Detecting traffic signs is an important factor to prevent accidents thanks to respecting traffic rules. However, building reliable traffic suitable for implementation on edge devices such as field-programmable gate arrays (FPGA) is a hard challenge. For this purpose, we proposed traffic signs detection based on convolutional neural networks. The YOLO object detection model was used as a detection framework. A lightweight backbone based on the squeezeNet model was proposed to achieve high performance with a light model that fits into the memory of the FPGA. Many optimization techniques have been applied to compress the model size such as pruning and quantization The proposed traffic sign dataset was implemented on the pynq z1 platform. training and evaluating the model were performed on the Chinese traffic sign detection (CTSD) dataset The proposed model has achieved an mAP of 96% and a processing speed of 16 FPS. The obtained results proved the efficiency of the proposed model in terms of speed and accuracy while being suitable for implementation on an edge device.