Coal India and AWS Innovation Challenge

This hackathon has a two-part demo of the working model and point of view on selected use cases. Please use the pre-defined template for your proposal submission.

Problem Statement 1

Develop a model which identifies the truck entering through the entry gate. Takes its snap shot and count. Identify if it is totally empty. If it is not empty, then raise an alert with meta info (time of entry, entry gate location, along with the video snippet) in JSON format. Similarly, when the truck is exiting, we need to take snap shot of the truck, count it as an exit truck, identify if it is loaded. In case truck is not loaded, generate an alert in JSON format with meta info (time, exit gate, and video snippet of the truck). 

Deliverables

1. Model with 80% plus accuracy. 
2. Participants need to gives us an approach paper/presentation which should must cover but not limited to following:
1. How is the model generalized? (Including but not limited to use of data augmentation etc.)
2. Type of model used; the basic architecture of the model. If a deep learning model is used, participants are expected to share and explain the base architecture and the level of transfer learning applied or if the model was trained from the scratch. We are looking at entire IP but a broader understanding. 
3. Event / Exception identification is a rare event. So, we need to understand how a rare event has been taken care of during as modeling and what metrics have been used. At a minimum, ROC curve, confusion metrics are required. 
4. What kind of hardware will be required for inference? 
5. Is there strategy for continuous learning proposed? We looking forward to all kind of innovation semi-supervised learning, online learning, reinforcement learning. 
6. How will the solution be scaled to multiple operating Opencast Mines?
7. Ways which will ensure more then 98% uptime of system with quality data collection being done continuously for desired inspection by CIL

Problem Statement 2

Number plate and vehicle registration number extraction from CCTV video feed. The objective shall be to detect the presence/absence of Vehicle Number Plate and extract the corresponding ROI. The solution should also detect, label, highlight and extract any/all occurrences of Vehicle registration number on a truck from the camera feed of the entry-exit gate. It is expected that an event will be generated in JSON format with meta info (time stamp, location id, and number plate). In case the number plate is not detected on the truck, a similar event will be generated in JSON format with meta info (time stamp, location id, and video snippet of the truck).

Deliverables

1. Model with at least 80% plus accuracy. 
2. Participants need to gives us an approach paper which should must cover but not limited to following: 
1. How is the model generalized? (Including but not limited to use of data augmentation etc.)
2. Type of model used; the basic architecture of the model. If a deep learning model is used, participants are expected to share and explain the base architecture and the level of transfer learning applied or if the model was trained from the scratch. We are looking at entire IP but a broader understanding. 
3. What kind of hardware will be required for inference? 
4. Is there strategy for continuous learning proposed? ? We looking forward to all kind of innovation semi-supervised learning, online learning, reinforcement learning.
5. How will the solution be scaled for approximately 180 operating Opencast Mines in multiple Areas/Coal Sidings?
6. Ways which will ensure more then 98% uptime of system with quality data collection being done continuously for desired inspection by CIL

Presentation on the point of view

In addition, to the demo and the approach of the model, we would require presentations on the proposed solution for two problem statements. 

1. Vehicle Re-ID for Tracing of Vehicle in Multi camera set up.
2. Boom barriers are to be in closed position and should open when Trucks are allowed in or out. 

Background

Coal India Limited (CIL) has approximately 180 operating Opencast Mines in 84 Area/Coal Sidings, Coal India has a huge operation involving, surface transportation of Coal using Trucks/Dumpers in every such mine/Area of CIL spreading across the country.

Considering the scale of operation, the task for managing such huge operations involving 5/6 depts. is considered as a challenge.

The basic Cyclic operation works in the following manner:

1. A Truck enters the Coal Loading area through the Entry Gate, through a boom barrier after Truck’s identity is authenticated by the RFID based authentication system.
2. Next it goes to the Weigh Bridge for recording the Tier weight, identification is through RFID.
3. Then the Truck goes to the designated Loading Area for Loading of Coal.
4. The loaded Truck goes to the Weigh Bridge for recording of Weight, identification is through RFID
5. The loaded Truck passes through the Exit Gate of the Coal Loading Area, through a boom barrier after Truck’s identity is authenticated by the RFID based authentication system.
6. The loaded truck may pass through an intermediate security Check point. Truck’s identity is authenticated by the RFID based authentication system.
7. Majority of the Trucks go to Railway Coal siding area for unloading of Coal, Truck’s identity is again authenticated by the RFID based authentication system.
8. The same cycle is repeated again from point 1.

It may be noted that there are 1000s of CC TV cameras installed in all important locations throughout CIL.

The problem is that such huge number of installed CC TV footages cannot be viewed continuously by any person for detection of every single exception types or any wrong doing/malpractices. As a result, such CC TV footage remains basically unattended/unsupervised and used/viewed post facto only when any incident is reported by security or other concerned depts.

To resolve this problem, an effective monitoring of the CC TV footage by developing an AI based solution can be considered as useful.

Camera Installed

Type: IP Fixed   Make  HIKVISION  Model DS-2CD2133-G0-1; DS-2CD132P-1; DS-2CD202PF-1;DS-42251W-AE;DS-2CD4A26FWD-IZ(h)(S).

MAKE: BOSCH Model: NT1-51022-A3S