In this work, we present a hybrid Oriented FAST and Rotated BRIEF (ORB) feature extraction pipeline that integrates the CPU, GPU, and NVIDIA Jetson’s Programmable Vision Accelerator (PVA) to optimize performance on embedded platforms. By assigning each stage of the pipeline to the most suitable processing unit, we reduce latency while preserving feature quality. Specifically, we replace the conventional Features From Accelerated Segment Test (FAST) detector with a Harris corner detector, achieving lower detection time and potentially enhanced localization accuracy. Experiments on Jetson AGX Xavier show that our hybrid design achieves a 1.16× speedup and 13% energy savings over the CPU+GPU setup, lowering energy use from 0.203 to 0.199 μJ/pixel. The pipeline remains SLAM-compatible and highlights the potential of heterogeneous scheduling for efficient, real-time visual processing on embedded systems.