Resource efficient Multi-task Post-processing Custom Hardware for CNN-based Real-time Road segmentation and Lane detection

In this paper, we implement acceleration hardware in post-processing for road segmentation and lane detection that minimizes resource consumption for custom design of convolutional neural network (CNN) based multi-task system. The proposed hardware is a low-cost and sequential architecture that good connectivity with inference accelerator, and the proposed optimization method can reduce memory consumption in the CNN-based segmentation system. The resource utilization was estimated on the UltraScale+ vu9p customized platform. The proposed post-processing hardware consumes within 2% of resources and accelerates our multi-task CNN with an inference accelerator.