Mathur, Ankit and Khattar, Saelig (2019) Real-time Wildlife Detection on Embedded Systems. Technical Report. Stanford InfoLab.
Modern machine learning algorithms are generally deployed on high performance GPUs, and require significant power and memory resources. However, such resources are not always available, especially when one wants to deploy such models on the ”edge.” In this project, we work with Jasper Ridge Biological Preserve in Stanford, CA, and develop an image classification system that can be deployed on the 18 camera traps they have setup around the preserve to detect wildlife. We experiment with and deploy various image classification models on a Raspberry Pi that can be connected to these camera traps. We show that our best model achieves a mean-per-class accuracy of 87.6% and can quickly run inference in real-time, on-device, using minimal power.
|Item Type:||Techreport (Technical Report)|
|Deposited By:||Andreas Paepcke|
|Deposited On:||30 Aug 2019 10:42|
|Last Modified:||30 Aug 2019 10:42|
Repository Staff Only: item control page