Edge video analytics demands vision models that operate under strict latency, energy, and bandwidth budgets while maintaining robust accuracy. This chapter surveys the artificial intelligence foundations that meet these constraints, tracing computer vision’s 60-year evolution from Roberts’ 1963 edge-extraction thesis through Marr’s representational hierarchy, the SIFT/HOG feature-engineering era, and the ImageNet-triggered deep-learning revolution. Object-detection families are contrasted among the two-stage R-CNN, fast R-CNN, faster R-CNN with RPN versus single-shot YOLO regressors that deliver real-time onboard inference. Generative adversarial networks are analyzed as synthetic-data engines for domain-adaptation and privacy-preserving augmentation, covering StyleGAN’s disentangled style-content space and CycleGAN’s unsupervised cross-domain translation under cycle-consistency losses. Finally, the chapter presents how vision transformers transplant self-attention and positional encoding from NLP to image patches, yielding global receptive fields without convolution and opening new compression-accuracy trade-offs for edge deployment.