In its tender, the state police defined the scope, accuracy and additional capabilities of the AI model and described a single app and web-based interface for the two software systems.
The Himachal Pradesh police are planning to use artificial intelligence and deep learning software to catch drivers who are wearing helmets. According to a tender viewed by MediaNama, the state police have tendered offers to develop two artificial intelligence (AI) -based video analysis software that can recognize drivers’ helmets and read their license plates. In addition to artificial intelligence, the system also requires provisions for the use of “deep learning” in software – deep learning technologies refer to artificial intelligence systems that learn for themselves over time by consuming data, interpreting patterns, and so on be able.
Why is it important? Artificial intelligence systems in policing or governance raise concerns about the degradation or expansion of the use of a product for purposes beyond the originally stated purpose.
The provision of deep learning creates the potential that this software (s) can be expanded to other functions and use cases in the future. The software also raises privacy concerns in the absence of a data protection law that can set purpose limitations and privacy rules.
In short: what the tender requires
The document states that the software will connect up to 10,000 CCTV cameras. It also sets some other specifications for the two products.
- For helmet recognition, the software should be able to recognize a helmet with a minimum size of ’40 x 40 pixels’ in the camera, be it on the driver or front passenger. This should have a detection rate of 95-100% during the day and be able to avoid pedestrians, rickshaws, etc.
- For license plate recognition, the software should be able to recognize different types of signs day and night. In addition, the software should also be able to recognize the license plate from a database and then look it up, generate warnings based on the status of the license plate, apply labels / tags to it (e.g. VIP, Offender, etc.) and ‘ Generate severity based on existing rules “. It is not clear which “rules” and “status” are referred to in the tender.
- In particular, Specifying the use of both software, the tender requires that events can be filtered by time range, location and attributes of people, vehicles and other objects such as visible colors, direction of movement and time of stay.
Detailed: Everything that the tender contains
For the license plate reading software
scope: The tender requires that software be based on “Artificial Intelligence and Deep Learning Models for Detection and Recognition”.
Types of license plates: According to the tender, this software must be able to recognize license plates of various types, such as:
- Not standardized
- Not reflective
accuracy: The tender requires 95-100% detection accuracy during the day and 85-100% during the night.
Additional ability: The tender also requires that the software is capable of:
1. Find the vehicle number plate from the database.
2. Generate an alert based on the status of the vehicle registration number in the database
3. Assign live label / tag to the recognized vehicle number plate (e.g. VIP / criminal etc.)
4. Generate the “severity” of the recognized license plate based on predefined rules. It is unclear what severity could mean here.
Filter ability: The software also asks for the ability to search vehicles by color, license plate, date and time, location and vehicle type.
For the helmet recognition software
Scope: The tender mentions the use of “deep learning” and “artificial intelligence” to detect two-wheelers without helmets.
Accuracy: The tendering requires an accuracy of 95-100% “during the day”
Reservations for detection: The offer requires that the software is capable of:
1. Recognize the driver and front passenger without a helmet. In the tender it is stated that this “must have a minimum size of 40 x 40 pixels in the field of view of the camera”.
2. Allow the user to define an area of interest for helmet injury detection.
3. Only recognize motorcyclists without helmets, avoid pedestrians, cyclists and rickshaw drivers as violators.
Software interface: what the information collected might look like
The tender describes a single app and a web-based interface for the two software systems. It also provides more information on the software’s artificial intelligence requirements.
1. This means a training tool that can annotate and label images to train new AI models and update the existing ones.
2. According to the tender, the training tool should also contain a list of all models available in the system that can be easily integrated into any app. This suggests that more than one AI model may be available or expected by the software.
1. They should be clearly configurable for each individual camera stream and have parameters for camera calibration, image quality improvement, night / day settings, etc.
2. You should be able to run on different cameras with different settings (eg different zones for break-ins, different lines for detecting line crossings, etc.) at different times of the day.
3. The software should have a “configuration page” suggesting any other AI models that may be available for the system, with clear specifications of their performance and hardware requirements.
- Provision of the video analysis application: Both software systems should be able to provide up to 3 video analysis applications simultaneously on a single camera. However, the tender does not specify what these video analytics applications would be.
What the data would look like
Live view functions: The app should enable a live view of the stream from every camera with “overlaid information on regions, objects, people and vehicles”.
Heat map and analysis: It should be able to provide an analytics dashboard with information on the “Pattern of Event” from different cameras in the form of a heat map.
Resource management view:
- The app should be able to provide a list of all the resources available in the system, such as computer servers, edge computing devices, and cameras.
- The status of each of the devices, whether online / offline, should also be available.
Event notifications: It should be able to deliver results in the form of events that include the screenshot with other metadata describing the event, e.g. B. detected objects, time stamp, camera / video.
App camera grid:
- The tender requires that the app has a “matrix for assigning, starting, stopping and scheduling each app on each camera”.
- The status of active and inactive apps is clearly visible through color-coded information.
Access to the software
Hosting: According to the tender, the app can be hosted in a “private cloud” or at the buyer’s premises.
user: According to the tender, both are software systems should manage up to 10,000 users in total and manage 1,000 users at the same time. However, the tender does not specify who these users will be or who will be allowed to use the software.
Access data and LAN: Although the tender does not mention who will be the users of the software systems, it states that the web interface “should be accessible on every system in the local area network (LAN) with access data”. HHowever, the same requirement is not mentioned for the app interface.
Other cases of automation in the police force in India
India has a long history of using automation in police and governance as this timeline shows:
July 2021: The Ministry of Railways has installed a facial recognition system (FRS) at 310 stations with plans to cover 673 more over time.
June 2021: Hyderabad Police received a legal notice for using FRS to detect lock violations.
April 2021: As part of the Smart Cities Mission, the government of Bihar has launched a tender for the use of a facial recognition system (FRS) linked to the Crime and Criminal Tracking Network & Systems (CCTNS) and other databases available from the police in Bhagalpur and Muzaffarpur.
January 2021: The Lucknow police had announced that it will install an FRS to alert the nearest police station if it detects a woman with a “troubled expression”.
November 2020: The Noida police were floating tender update its traffic management system. That tender called for nearly 1,000 cameras to be installed across the city, including surveillance cameras and separate license plate recognition cameras.
May 2020: The Ministry of Housing and Urban Development announced that quarantine stations in at least four hospitals and one hotel in Uttarakhands Dehradun are being monitored around the clock with an FRS-enabled CCTV surveillance system.
April 2020: Hyderabad Police used artificial intelligence over 2,000 CCTV devices to identify mask violators.
Do you have anything to add? Subscribe to MediaNama and post your comment