5.3 A Data-Compressive 1.5b/2.75b Log-Gradient QVGA Image Sensor with Multi-Scale Readout for Always-On Object Detection

Abstract

Histograms of Oriented Gradients (HOG) are attractive features for object detection in embedded vision applications, as they provide a good trade-off between complexity and detection accuracy. A custom 8b CMOS imager that computes these features on-chip consumes only 52pJ/pixel [1]. However, a complete system also requires a backend detection algorithm, which consumes 940pJ/pixel in an optimized implementation [2]. As shown in the system study of [3], this imbalance mostly stems from the large amount of data seen by the detector. To remedy this issue, the work of [3] studies a feature-extraction approach that aggressively log-quantizes the data, thereby eliminating unnecessary illumination-related bits from the histograms. The custom log-gradient image sensor described in this paper demonstrates this concept in CMOS. It consumes 127pJ/pixel and offers two log-gradient modes (1.5b and 2.75b) along with multi-scale readout. With several algorithmic enhancements enabled by the log gradients as described in [3], the resulting HOG feature compression ratios are 25× (1.5b) and 9.5× (2.75b) when compared to a standard 8b image (without image pyramid). Using 1.5b log gradients, [3] conservatively estimates a 3.3× reduction in backend detection energy for a deformable parts model (DPM) based detector.