Duration: Self-paced 21 days (Recommended learning hrs: 6-hrs/week)

Dates: 28 Sept, 2020 to 19 Oct, 2020

Registration Deadline: 26th September 2020

Registration LinK: https://iioe.ichei.org/training  (Enrolment/Registration guide is attached)

Leading Instructor: Dr. Wen Zheng, Assistant Professor of the Department of Communications and Computer Engineering at Waseda University, Japan

Prospective Learners: Educators and professionals in higher education who wish to enhance data literacy and data analytic skills in their teaching and research

Recommended devices:  Participate in IIOE Training via desktop computers and laptops.

On Tuesday, September 15, 2020, NCAI associated Lab AGVRL_KICS_UETL organized a successful webinar by the name of ‘on Node deep Learning’, under the supervision of Dr. Usman Ghani Khan (PI_NCAI). A number of participants including NCAI associated labs i.e. AGVRL, CVML, and private software houses attended the webinar.

The key speaker of the webinar was Muhammad Ahmad Hassan (sr. research officer) AGVRL lab. In this session, he very clearly enlightens the concept of current paradigms of security and surveillance system, problems associated with traditional manual and automatic surveillance systems, current challenges to the CCTV security systems. Further, he also proposed the practical, applicable and reliable solution to all issues related to security and surveillance systems in on node deep learning.

Furthermore, in explanation with the current scenario, he said; our security systems are being shifted to cameras, thermo graphics cameras; video streams based authorization and detection, person authorization, object detection, and site surveillance. But here, the problem is, all this installation needs a force of human resources to monitor and process all the streamed data. That is quite a time-taking and of course, we cannot ignore the chances of human error.     

Here, he quoted the example of a safe city. There are almost ten thousand cameras that are continuously generating streams. Let’s if we apply the same concept of “Safe city” to all the provinces, it would be the required installation of almost fifty thousand cameras. These cameras would continuously generate live streams and producing tons of data. If this data is to be filtered, monitored, and processed by a human, it would take approximately ten thousand manpower and months of time. But still, after doing all that, human error and negligence cannot be ignored. As humans cannot pay attention consistently for a long period, he has desires and emotions too. These are unavoidable limitations of traditional security and surveillance systems.

Further limitations of hybrid security surveillance systems are;

1. Heavy servers required for video processing.
2. Central server dependency.
3. On Node decision.
4. Network delay.

In case of offline processing the limitations are;

1. Difficult to retain such Mammoth data.
2. Processing will incur resources.
3. Financial constraints.
4. Technical constraints.
5. Spatial constraints.

Solution for current surveillance challenges

Mr. Muhammad Ahmad Hassan proposed an authentic, reliable solution to all the above-mentioned problems. He also demonstrates the practical application in real time. That was noted and appreciated by all the attendees. According to his proposed solution,

1. If each node in the surveillance system work, process and make the decision individually then the system can easily tackle the above-mentioned challenges.
2. Processing and decision should be on the individual node and final results can easily be transmitted to the main server.

• Deep learning models can also be implemented if financially affordable. This can easily solve our system problems like; data storage, network delay and on node decision etc.

1. On Node deep learning-based automation can also help us to overcome all such problems. For this, we need to embed deep learning models on node.

Further, he also discusses in-depth about on Node processing devices. Some of them are;

Raspberry Pi4

FPGA (Sparten 3e)

NVidia Jetson  Nano

Mr. Muhammad Ahmad Hassan also discussed in detail the working model, functionality, and useful features of TensorRT. I.e. it provides high-performance inference in deep learning. It has low latency and high throughput. Its libraries are available in C++ and python. Most importantly, it is compatible with almost all the available framework of deep learning like;

1. Tensorflow
2. Keras
3. MxNet
4. Pytorch
5. Caffe etc.

The webinar was concluded with a number of live demos and answering the questions raised by the audience. Speaker tried to satisfy them all with comprehensive answers and examples.  All attendee was thankful of the institute for arranging such an advance and knowledge-rich webinar.