Presentations from the AI conference held by EiTESAL that showcases the challenges met in the AI market in Egypt. presented by eng. Ahmed Khalil from Synaplexus
5. Unauthorized areas are marked by an imaginary line in the
field of view. When the line is crossed/tripped the detection
event is triggered
5
Unauthorized Access /Trip Wire Detection
6. For example a luggage is left at a train station by a person.
6
Unattended Object Detection
7. A person remains in a monitored area for certain time where
loitering is not allowed
7
Loitering Detection
8. Anomaly detection can be defined as the task of detecting
unusual patterns of behavior such as cycling in a pedestrian
area.
8
Anomaly Detection
9. The level of Crowdedness in a given scene i.e. determining
whether a given area is lightly , moderately or highly
packed
9
Congestion Detection
11. ● Object Classification can be defined as the task of
identifying whether a certain object of known class is
present in an image
11
Object Classification
12. ● The object detection can be defined as the task of finding
a bounding box that surrounds an object of known class in
an image
12
Object Detection
13. ● The objective of the segmentation is to label each pixel of a
given image
● The labels belongs to a set of known classes
13
Segmentation
14. ● The Objective of the density estimation task is to determine
relative crowdedness of a given area.
14
Density Estimation and Head Counting
15. ● Face recognition task has the objective of detecting the
face region and identifying whether the face is known or
not
15
Face Recognition
22. 22
Vision at the Edge
reduced or eliminated
is eliminated
Cloud analytics is charged per camera per month.
usage is extremely efficient
The usage of cloud analytics requires a high speed high
capacity network to send the videos efficiently to the cloud.
Running the analytics locally on the camera not on the cloud
eliminates the Cloud Overhead which is consumed by:
• Sending the video feeds to the cloud.
• Running the video analytics on the cloud.
• Take proper action whether to trigger an alarm locally or send
notification to cell phone
23. Challenges:
Deep learning algorithms are mostly designed with relaxed
requirements for computational power and memory
Scarcity of Edge HW supporting existing deep learning
frameworks
Fusing shared data between the local Cams requires
modification to readily trained models
23
Vision at the Edge