Fast Traffic Sign and Light Detection using Deep Learning for Automotive Applications

Traffic sign and light detections are core components of Advanced Driver Assistance Systems (ADAS) and self-driving vehicles. To this end, the automotive industry is widely exploiting computer vision (CV) and deep learning (DL) techniques. This paper presents a lightweight traffic sign and light detector by harnessing a single-stage, single-shot multi-object detector (SSD). For accelerating the inference speed of the detector, its original backbone, VGG16 is replaced by MobileNet V2 that expertly manages detection speed and network size. In autonomous driving, quicker detection performance with respect to the distance of an object is of particular interest, for a comfortable braking. However, farther distance makes the objects to be detected appear smaller. Unfortunately, the original SSD struggles to detect small objects. Thus, this work further optimizes the number of feature map layers of the SSD for the detection of small objects along with a better trade-off between detection precision and inference time. Experimental analysis confirms the effectiveness of the proposed model, which achieves 2 times (or more) faster detection time than the baseline SSD models and a competitive precision of 76.7%.