ORBCam: In-Sensor ORB Feature Processing for Ultra-Low-Power Visual-Inertial Odometry
Abstract
In visual–inertial odometry (VIO) systems, image readout and data movement between sensor and processor are increasingly recognized as the dominant power bottleneck, overshadowing on-chip computation. To address this, we present ORBCam, a cross-layer sensor–estimator co-design that eliminates image readout and directly generates motion-required feature measurements within the sensor subsystem. Instead of exporting images or descriptors, ORBCam transmits only quantized pixel coordinates and flow measurements to the host. In system-level simulations at 752 × 480 resolution and 100 FPS, ORBCam is compared against a conventional image sensor consuming 7.88 mW for full-frame acquisition and transmission. ORBCam reduces sensing power to 0.59 mW, achieving up to 13.3× energy efficiency improvement while maintaining comparable odometry accuracy.