Deep learning is becoming increasingly popular for visual understanding use cases on edge devices, such as image classification and object detection, and there has been an increasing demand for running basic computer vision and deep learning models on edge devices due to concerns about privacy and security.
There are several implementations of SSD with deep neural nets, such as GoogLeNet and VGG, which provide low-throughput object-detection solutions on edge devices. Srinivasa Karlapalem demonstrates a new SSD network with SqueezeNet for high-throughput single-shot multibox object detection (SSD) on edge devices using FPGAs, specifically for surveillance. This network is used for multiclass object detection on IoT platforms such as Apollo Lake and SkyLake, accelerated with Arria 10 FPGAs.
The solution uses live and prerecorded video streams or the Pascal VOC dataset as input. Frames are extracted and resized to be fed into the Squeezenet SSD network. The SSD network uses pretrained models and input frames to determine bounding boxes of objects and their classes. Srinivasa details the detection throughput (images/sec) and other metrics such as CPU and FPGA utilization.
Srinivasa Manohar Karlapalem is a software engineer at Intel, where he develops and optimizes novel deep learning-based applications on compute and memory-constrained edge and IoT hardware architectures. He has extensive experience optimizing programming language VM runtimes and OS stacks on modern, high-performance Intel architectures. He holds an MS in computer science from the Georgia Institute of Technology.
©2018, O'Reilly Media, Inc. • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • firstname.lastname@example.org