This paper presents a novel traffic sign detection (TSD) approach using off-the-shelf onboard vehicular cameras. Assuming that the camera intrinsic parameters are obtained offline, an online calibration scheme is used to estimate the extrinsic camera parameters, and Regions of Interest (ROIs) are created in the image domain based on the expected geometry and location of the traffic signs. Within these ROIs, the scale variation of the sign and background complexity are limited, allowing the development of lightweight Convolutional Neural Networks (CNNs) for TSD. Our experimental results for Brazilian traffic signs indicate that the proposed approach presents better accuracy than state-of-the-art methods at faster running times, being 62× faster than lightweight models such as YOLOv4-tiny using a Raspberry Pi 3 hardware. Data, code, and examples of processed videos are available at https://github.com/maubrapa/FTSD_DOC/.