FPGA-Accelerated Semantic Segmentation for Urban Scenes

To solve the problem in urban scene semantic segmentation due to complex conditions and vague boundaries, an FPGA-based mining area semantic segmentation algorithm was designed. The model follows an encoder-decoder framework, with ResNet being the encoder and Semantic FPN being the decoder. To accelerate and deploy the FPGA platform, Li norm pruning method was used to compress the ResNet model, and Vitis AI tool to quantify and compile the pruned model. The model can achieve a frame rate of 17.6 FPS on the Xilinx Kria KV260 platform with a power consumption of 6.6W, and the energy efficiency ratio is significantly higher than that on the CPU or GPU platform.