Analysis of performance tradeoffs for embedded HOG feature extraction
- Recent mobile vision applications demand energy-efficient real-time object detection. Specialized hardware design is needed to push the limits of both performance and energy-efficiency. While such hardware has been demonstrated for backend detection, current imager frontends consume a significant fraction of total system energy. Therefore, additional system-level energy savings may be achieved by reducing the energy requirements of frontend image capture. At the same time, it is crucial that the energy saving techniques used do not significantly degrade object-detection performance. This dissertation studies the effects of frontend imager parameters on object detection performance and energy consumption. A simulation framework, including a largescale RAW image database for object detection, is developed. And simulation results quantifying the tradeoff between pixel bitdepth and HOG-based object detection performance are presented. A custom version of HOG features based on 2-bit pixel ratios is introduced, and shown to achieve superior object detection performance for the same estimated energy compared to conventional HOG features. A frontend hardware implementation capable of extracting these features at multiple scales is proposed, and a system-level energy analysis is performed. This energy analysis suggests a potential 19X reduction in I/O energy and 3.3X reduction in backend detection energy compared to conventional object detection pipelines.
|Type of resource
|electronic; electronic resource; remote
|1 online resource.
|Stanford University, Department of Electrical Engineering.
|Dutton, Robert W
|Dutton, Robert W
|Statement of responsibility
|Submitted to the Department of Electrical Engineering.
|Thesis (Ph.D.)--Stanford University, 2017.
- © 2017 by Alexander Baktosh Omid-Zohoor
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...