Real-Time Object Detection On Low Power Embedded Platforms
| Autor: | Harikrishna Muralidhara, Srinivas Kruthiventi S S, Aashish Kumar, Sambuddha Saha, George Jose |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
Artificial neural network
Computer science business.industry Quantization (signal processing) Deep learning 05 social sciences Wearable computer Advanced driver assistance systems 010501 environmental sciences 01 natural sciences Object detection Embedded system 0502 economics and business System on a chip Artificial intelligence 050207 economics business 0105 earth and related environmental sciences |
| Zdroj: | ICCV Workshops |
| DOI: | 10.1109/iccvw.2019.00304 |
| Popis: | Low power real-time object detection is an interesting application in deep learning with applications in smart wearables, Advanced Driver Assistance Systems (ADAS), drone surveillance systems, etc. In this paper, we discuss the limitations with existing networks and enumerate the various factors to keep in mind while designing neural networks for a target hardware. Based on our experience of working with TI embedded platform, we provide a systematic approach for designing real time object detection networks on low power embedded platforms. First stage involves identifying the optimal layers for the hardware, by understanding it's computational and memory limitations. The next step is to use these layers to come up with a basic building block that has low computational complexity. The final stage involves using model compression techniques like sparsification/quantization to accelerate the inference process. Based on this design approach, we were able to come up with a low latency object detection model HX-LPNet that operates at 22 FPS on low power TDA2PX System on Chip(SoC) provided by Texas Instruments (TI) |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|