Video analytics has been evolving into a cross-layer discipline that orchestrates canonical stages—capture, encoding, decoding, preprocessing, segmentation, detection, recognition, tracking, and multisource fusion—while allowing CNN backbones to merge several phases into one forward pass for edge efficiency. Its performance can be evaluated by a multidimensional metric, including precision, recall, F1 score, end-to-end latency, computation and feedback delays, network usage ratios, energy consumption, etc. These metrics shape divergent real-world deployments, city-wide surveillance, behavior anomaly, intelligent-transportation cameras, medical analytics in MRI, CT, PET, and AR/VR headsets, etc. Owing to the substantial computing resource required, the video analytics paradigm has to exploit the cloud-scale training, edge-level inference, and end-device sensing into a single continuum whose configuration knobs—model depth, quantization, frame resolution, key-frame interval, offload ratio, and federation strategy-are tuned to balance accuracy, latency, bandwidth, and energy for each application domain.