Object-adaptive CMOS Image Sensor with Embedded Feature Extraction Algorithm


Distributed sensor nodes typically operate under the constraint of limited energy source, and power consumption is an important factor to extend the lifetime of sensor systems. Several low-power imagers have been reported for the application of wireless sensor network. However, the biggest power consumption comes from wireless signal transmission due to the large bandwidth of image signals. One way to reduce the bandwidth is to generate signals only when event happens by monitoring temporal changes. However, this event-based imaging has extraneous redundancy because the sensor may also respond to environmental conditions, such as change of illumination or background movement in addition to actual target objects.

In this work, we report motion-triggered object-of-interest (OOI) imaging to suppress the redundancy in imaging as well as transmission of signals. In most of time the sensor is in sleep mode until it is triggered by motion. When it wakes up, it generates and transmits 8-b features for object detection. The signal processing unit which resides either in the host or in the sensor node, performs detection of objects and feeds back 1-b request signal to initiate further imaging operation if the object-of-interest is identified. Among many feature extraction algorithms, we incorporated the histogram-of-oriented-gradients (HOG) because it gives a high detection rate of objects in simple operation. The HOG feature output requires only 3.5% of bandwidth when compared with conventional 8-b image capturing. In this work, we implemented the HOG feature extraction algorithm using mixed-signal circuitry in order to save both power and area.

A prototype chip has been fabricated using 0.18μm CMOS process. In order to verify the performance of the integrated feature extraction unit, we tested the object detection from the extracted features using 200 pedestrian images from DaimlerChrysler dataset. The test result shows 94.5 % detection rate. We achieved a normalized power of 13.46 pW/frame•pixel in motion sensing and 51.94 pW/frame•pixel in feature extraction.


Related publication
  1. J. Choi, S. Park, J.Cho, E. Yoon, "A 3.4 μW CMOS Image Sensor with Embedded Feature Extraction Algorithm for Motion-Triggered Object-of-Interest Imaging," Technical Digest of International Solid-State Circuits Conference, pp.478-479. Feb. 2013.