Efficient Real-Time Computer Vision on Embedded Systems with Neural Cellular Automata

Neural Cellular Automata (NCA) demonstrate competitive performance with hierarchical Convolutional Neural Networks (CNN) for real-time computer vision tasks on embedded systems and resource-constrained edge devices. The proposed architecture, LNCA (Latent Neural Cellular Automata), was developed through systematic ablation of 14 design dimensions, achieving 3.7 ms inference latency at 270 frames per second (FPS) on a Moto G Stylus 5G 2023 smartphone with a model size of 55 KB. LNCA achieves 3× lower latency than BiSeNet V2 and 8-89× smaller model size than comparable CNN segmentation architectures while maintaining comparable Intersection over Union (IoU). Ablation studies reveal that depthwise convolutions for both perception and encoding outperform other design options. NCAs provide a practical foundation for real-time computer vision on edge devices.