Utilizing AI & GPUs to Build Cloud-based Real-Time Video Event Detection Solutions Zvika Ashani CTO.

Slides:



Advertisements
Similar presentations
Applications of one-class classification
Advertisements

SLA-Oriented Resource Provisioning for Cloud Computing
A new Network Concept for transporting and storing digital video…………
XProtect® Expert 2013 Product presentation
A Java Architecture for the Internet of Things Noel Poore, Architect Pete St. Pierre, Product Manager Java Platform Group, Internet of Things September.
Axis Intelligent Video Intelligence where you need it.
Smart Cities VMS This is a Smart Cities inspired software platform that centres around the surveillance cameras (CCTV) and that can integrate other IoT.
Jordan Wissel Eric Lewis Sarah Basile. Introduction This presentation will analyze: Overview/History Implementation Advantages/Disadvantages Security.
Net Optics Confidential and Proprietary Net Optics appTap Intelligent Access and Monitoring Architecture Solutions.
Axis Intelligent Video Strategy Focus on Embedded Video Analytics.
Efficient Visual Object Tracking with Online Nearest Neighbor Classifier Many slides adapt from Steve Gu.
Transforming video & photo collections into valuable resources John Waugaman President - Tygart Technology, Inc.
Enterprise Cloud Computing
Flexible intelligent solid video management software.
© 2014 IBM Corporation Does your Cloud have a Silver Lining ? The adoption of Cloud in Grid Operations of Electric Distribution Utilities Kieran McLoughlin.
August Video Management Software ViconNet Enterprise Video Management Software Hybrid DVR Kollector Strike Kollector Force Plug & Play NVR HDExpress.
Cloud Computing from a Developer’s Perspective Shlomo Swidler CTO & Founder mydrifts.com 25 January 2009.
Company Confidential Konnet ViAn Konnet ViAn P. Ltd. 102, Shiv Shakti Compex Baner road, Pune , India /5 Computing.
Flame & Smoke Detection System Flame & Smoke Vision Detection is an intelligent vision-based analytics system which can timely detect suspicious fire or.
MAXPRO NVR HONEYWELL - CONFIDENTIAL File Number MAXPRO NVR 2.5 HON Cameras Support – HDZ PTZ, ONVIF cameras (equivalent ONVIF models to current.
SUSE Linux Enterprise Server for SAP Applications
Comparing TensorFlow Deep Learning Performance Using CPUs, GPUs, Local PCs and Cloud Pace University, Research Day, May 5, 2017 John Lawrence, Jonas Malmsten,
Effects of Limiting Numerical Precision on Neural Networks
WHY VIDEO SURVELLIANCE
WHY VIDEO SURVELLIANCE
Connected Infrastructure
When deep learning meets object detection: Introduction to two technologies: SSD and YOLO Wenchi Ma.
REAL-TIME DETECTOR FOR UNUSUAL BEHAVIOR
Deep Learning Software: TensorFlow
Cloud Technology and the NGS Steve Thorn Edinburgh University (Matteo Turilli, Oxford University)‏ Presented by David Fergusson.
ICON Signals Event Alert Notification Platform Overview March 28, 2017
Connected Living Connected Living What to look for Architecture
How Cutting Edge Big Data and Analytics Lets J. D
Benchmarking Deep Learning Inference
Big Data A Quick Review on Analytical Tools
Microsoft Operations Management Suite Insight and Analytics
Chilimbi, et al. (2014) Microsoft Research
Cloud Computing & ANalytics
Chapter 19 Cloud Computing for Multimedia Services
Connected Living Connected Living What to look for Architecture
Connected Infrastructure
M.Tech Technical seminar Presentation
Remote Monitoring solution
Exam VCE Questions
Erik de Nooij, IT Chapter Lead Fraud&Cybersec.
Wavestore Integrates…
Object Recognition & Detection
Layer-wise Performance Bottleneck Analysis of Deep Neural Networks
Voice Analytics on Microsoft Azure Allows Various Customers to Get the Most Out of Conversations with Clients Through Efficient Content Analysis MICROSOFT.
Wavestore Integrates…
AKAMAI INTELLIGENT PLATFORM™
20409A 7: Installing and Configuring System Center 2012 R2 Virtual Machine Manager Module 7 Installing and Configuring System Center 2012 R2 Virtual.
Embedded & Cloud Platforms
What's New in eCognition 9
The Pitfalls and Guidelines for Mobile ML Benchmarking
Technical Capabilities
TrinityIoT Premises Monitoring.
Nenad Stefanovic and Danijela Milosevic
WHY VIDEO SURVELLIANCE
WHY VIDEO SURVELLIANCE
The Performance of Big Data Workloads in Cloud Datacenters
AI-Powered Information Governance
Heterogeneous convolutional neural networks for visual recognition
How Edge Intelligence Accelerates Network- Compute-Convergence (NCC)
Coventry University, UK
What's New in eCognition 9
Utilizing the Network Edge
What's New in eCognition 9
Microsoft Virtual Academy
Introducing MagicInfo 6
Presentation transcript:

Utilizing AI & GPUs to Build Cloud-based Real-Time Video Event Detection Solutions Zvika Ashani CTO

Background Agent Vi has been developing video analytics solutions for surveillance applications for the past 10 years Our products provide a diverse set of capabilities such as real time alerts, forensic search and statistical analysis Our enterprise solution is installed on premise using customer provided servers 2 years ago we set out to create a SaaS platform that will provide real time video analytics capabilities to our customers © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

What’s the big deal with VA in the cloud? Deep Learning for object detection and classification – a de-facto standard Image analytics as a service – numerous providers out there (AWS, Azure, Google and others). Applications such as image tagging, facial recognition, image moderation and more. Offline (batch processing) of video is also starting to become available from these same providers. Mostly used for video indexing and search. Real time video analysis is still a challenge, mainly due to bandwidth constraints © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Platform requirements Easy to deploy, does not require any software installation on site Analyzes surveillance video and provides real time alerts based on user defined rules Provide state of the art probability of detection and false alarm rates Support any IP surveillance camera that can provide an RTSP stream Utilize low bandwidth per camera (less than 50kbps) Use low cost HW on site Operate at near real time (latency should not exceed 200ms) Operational cost per camera per month should be very low © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Challenges Cannot upload a high quality video stream to the cloud due to BW limitations Cannot deploy a GPU to the user sites due to cost considerations Solution Use a distributed processing architecture Perform initial processing of the video stream at the edge to reduce the amount of BW required Send the results of the initial processing to the service in the cloud Use a combination of CPU’s and GPU’s in the cloud to complete the processing and detect events © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

All communication is encrypted using SSL Service Architecture Agent Vi Support Internet 30-50 kbps per camera InnoVi Edge Customer Internet Feature streams Customer InnoVi Edge Monitoring Stations Internet Admin All communication is encrypted using SSL © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

At the edge In the cloud Video decoding Foreground detection Blob segmentation Image cropping 50 kbps upstream 10 kbps downstream In the cloud Object detection and tracking Object classification Rules processing and event generation Other tasks such as automated scene learning © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Our Neural Network implementation Mostly use CNN’s for object detection and classification and LSTMs for other tasks Classify images into one of 15 classes Images are B&W 64x64 pixels Publicly available pre-trained models on ImageNet (or other data sets) are irrelevant due to differing image format and fields of view Need to balance network complexity between accuracy and cost per camera Use the Caffe framework for implementing the various network topologies © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Only real surveillance footage is used - people © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Only real surveillance footage is used - vehicles © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Only real surveillance footage is used - noise © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Analytics Server - Initial implementation Customer Analytics Server Internet InnoVi Edge N servers with CPU + GPU Assumption – there is an sufficient number of servers available on AWS to allow the solution to scale to support any number of cameras Runs both CPU intensive and GPU intensive tasks Was actually CPU bound, average GPU utilization was about 25% © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Analytics Server - Micro service implementation Customer Internet InnoVi Edge Analytics Server Analytics Service N servers with CPU Classification Service M servers with CPU + GPU Split analytics server to GPU and CPU tasks Each server cluster scales separately Over 90% utilization achieved on both clusters © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Classifier Service Benchmark g2.2xlarge – NVIDIA GRID GPU $0.7 per hour g3.4xlarge – NVIDIA Tesla M60 GPU $1.21 per hour © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Examples © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary

Cost can be further decreased Next obvious step is implementing inference with INT8. AWS has yet to announce support for EC2 instances with INT8 support Cost per image is expected to go down by a factor of 2x – 4x © 2017 Agent Video Intelligence Ltd. Confidential and Proprietary