Deep Learning At the edge of the network Hugo Latapie / Enzo Fenoglio Santosh Pandey, JP Vasseur, Rob Liston, Dan Tan, Xiaoqing Zhu, John Apostolopoulos with the newly formed IoE Video Analytics virtual team
HPC Since 2011 • Audio/video metadata extraction • Multi-modal TV UX • Encrypted Network Traffic Classification • Video compression
SLN Architecture Threat Grid, ISE OpennDNS, WBRS, ... Other • Orchestration of Distributed Learning TI feeds SCA Agents (DLAs) Threat • Intel Advanced Visualization of anomalies • SCA Centralized policy for mitigation • Interaction with other security components Controller such as ISE and Threat Intelligence Feeds Internet • North bound API to SIEM/Database (e.g. Splunk) using CEF/CIM format Public/Private • Evaluation of anomaly relevancy Internet • Sensing (knowledge) : granular data collection with knowledge extraction from NetFlow but also Deep Packet Inspection DLA DLA on control and data plane & local states DLA • Machine Learning : real-time embedded Branch 2 behavioral modeling and anomaly detection • Control : autonomous embedded control, advanced networking control (police, Branch 1 shaper, recoloring, redirect, ...)
Evolution of location use-cases 93% • Healthcare Asset management & wayfinding • Retail Engage shopper in aisle & deliver proximity-based offers • Museum Enabling the Digital docent • Office Wayfinding & workspace optimization
Technology – network based location • State of the art: • 5-7m accuracy • Multilateration on WiFi Client based on RSSI at multiple APs • Cisco-Hyperlocation: • 1m accuracy • Increase accuracy & reliability • AP connected clients • Not applicable to non-connected/probing clients or tags • Add Angle-of-Arrival in addition to RSSI • Introduce enhanced location with 1AP - RSSI Receive Signal Strength Indicator - ToF Time of Flight - DRTT Differential Round-trip-time
How Hyperlocation Works Built on Cisco Unified Access HyperLocation Access Points Controller MSE (Virtual/Physical) (Virtual/Physical) Depending on Application Layer CMX Analytics CMX Connect LOCATION DATA APPLICATION DATA Mobile Application Server Analytics UI
MVP 1 -- Counting
Scalable video analytics in crowded scenes Phase 1 Phase 2 • General DL pipeline based on nvidia-docker overlay networking and swarm with support for thousands of nodes and 10’s of thousands of containers • High performance video pipeline based on gstreamer
Tracking Overview • Tracking involves extracting the spatio-temporal history of each person in a scene solving the assignment problem using Jonker- Volgenant algorithm • We use a Convolutional Network (CNN) + LSTM • Tracking estimation is performed by multiple Kalman filters assigned to each tracklet Localization Input RGB Tracking CNN/ Estimation frames estimation KF/JV LSTM
Back to data fusion
Machine Learning from End Point to Cloud An intersection of technology and trends: • Recent camera sensor technologies offer advanced capabilities • Advanced cameras + analytics create powerful IoT sensors • Visual analytics and data fusion fit naturally into the fog architecture The industry is moving towards scalable, flexible, distributed analytics platforms Advanced Cameras Heterogeneous Compute Platform CPU + GPU/FPGA IoT Cloud Software
2 innovative pilots proposed by the City: outdoor/indoor 1. Better Understanding of the public space CDP, MERAKI wifi, CMX, LoRaWAN, cameras, video analytics, Energy / Asset Management 2. Optimization of energy management in public buildings
Cisco Data Center Portfolio and NVIDIA • Virtual workstation for high-end graphics applications Cisco UCS C240 M4 Rack servers • • Support NVIDIA GRID 1.0 and 2.0 with K1/K2 and M60 cards • Support for Magma ExpressBox for higher density • Cisco UCS B200 M4 Blade servers Recently introduced NVIDIA M6 MXM support on Blade server • Cisco HyperFlex – 2 nd generation HyperConverged • platform Phase 1: K1/K2 support • • Phase 2: M6/M60 support • Deep Learning and HPC Cisco UCS C240 Rack servers with TESLA K80 •
Recommend
More recommend
