Vehicle Windshield Detection by Fast and Compact Encoder-Decoder FCN Architecture

Vehicle semantic segmentation is critical in many advanced driving assistance systems, traffic management, and security surveillance systems. Most of such systems are deployed on low computational embedded systems located in the vehicles or in remote gantry and roadside poles. While fully convolutional networks have been proved to be a powerful classifier being able to make inference on every single pixel of the input image, they entail high computational costs even for the inference process. In this paper, a vehicle windshield semantic segmentation is proposed utilizing a fast and compact encoder-decoder architecture of a fully convolutional network implemented in a low-power embedded system. The performed qualitative and quantitative performance measurements exemplify a real-time portable embedded solution which is competitive in terms of performance and inference time.