Real-time object detection refers to models that can locate and classify objects in images or video frames fast enough for live applications, typically at 30 or more frames per second on the target hardware. The YOLO (You Only Look Once) family has defined this category since 2015 by treating detection as a single-pass regression problem rather than a multi-stage pipeline. From the original YOLOv1 through YOLOv8, YOLO11, and YOLO26, each generation has pushed the speed-accuracy boundary further.

Transformer-based detectors have recently entered the real-time space. RT-DETR (Baidu, 2023) and D-FINE achieve competitive accuracy and throughput while eliminating NMS post-processing entirely, using one-to-one label assignment during training. Key techniques for hitting real-time speeds include efficient backbone design (CSPNet, EfficientRep, MobileNet), depthwise separable convolutions, multi-scale feature fusion (FPN/PANet/BiFPN), model quantization (FP16/INT8), and hardware-specific compilation (TensorRT for NVIDIA, CoreML for Apple, LiteRT for ARM).

Deployment targets range from cloud GPUs (NVIDIA T4, A10) to edge devices (Jetson Orin, Raspberry Pi, mobile phones). Real-time detection powers live video surveillance, autonomous driving perception, industrial quality inspection on production lines, augmented reality overlays, and robotic pick-and-place systems.