CrimeWatch.AI

Inspiration

In San Francisco, security remains a formidable challenge, marked by alarmingly high crime rates. Despite escalating law enforcement budgets, crime persists at concerning levels, presenting a significant strain on resources. With traditional surveillance methods proving inefficient and manpower-intensive, the need for a cost-effective solution is urgent. In 2023 alone, the city recorded 53 homicides, 31,428 larceny thefts, 6,571 motor vehicle thefts, and 2,693 robberies, underscoring the gravity of the situation. Yet, traditional CCTV systems, prone to human error and passive monitoring, have failed to deliver timely responses, leaving security compromised.

What it does

Our solution is an Gemini-powered CCTV software and community platform that revolutionizes urban surveillance. By leveraging cutting-edge artificial intelligence algorithms, our software automates the monitoring and analysis of CCTV footage, reducing the need for manpower-intensive surveillance operations. This not only enhances the effectiveness of security measures but also significantly reduces costs associated with traditional surveillance methods. Our software provides real-time alerts for potential security threats, enabling swift response and intervention.

Beyond that, our software aims to not only allow for individual isolated businesses to aggregate information of the crimes happening to them, but also allow cross sharing of such reports and information with nearby communities, as crime is often conducted by a small set of individuals within the same community. By automating the crime detection and report generation with gemini, it becomes easy to build upon it with the next step tp share information and promote a stronger security based mindset within the community as well.

By making security and surveillance easier and more affordable, our solution addresses the pressing challenges faced by cities like San Francisco, ultimately fostering safer communities.

How we built it

We used React and Typescript for our frontend, and made use of multiple Gemini 1.5 pro calls in the backend in order to get 1) Analysis of whether a 10 second video clip is associated with a crime 2) Summary of the suspicious activity 3) A report of the suspicious activities happening in the area within 7 days of the crime occurring. We made use of fastapi to build our backend APIs and hosted it on azure during testing

Challenges we ran into

Due to the nature of the CCTV footage being crime related, there were times when the responses would be blocked by the Gemini's safety filters, hence we had to disable these filters to allow it to work in a wider range of cases.
We aimed to train a filtering model different model specifically to allow crime related outputs, but filter out other types of harmful content, however due to time constraints we were unable to complete it for deployment.

Accomplishments that we're proud of

Buiding a product that is not only able to be used for a small use case, but that can have massive ripple effects, by utilising the capabilities of Gemini 1.5 pro to massively automate manual tasks. Furthermore, this also unlocks the potential for a community building, adding massively to our value propostion.
Prompting gemini to give consistent and constrained outputs

What we learned

Gained a better understanding of the capabiliities of gemini as a multimodal model, utilising that to build our product

What's next for CrimeWatch.AI

With our focus on community building, we would like to increase the functionalities to allow cross sharing of information, and allowing multiple platforms to share and update information in real time, such as through the use of an interactive map with which users can use to retrieve and view reports of nearby crimes, as well as suggestions to improve security
Once we are able to build up our community and gain more data, we will also be able to start recognising and detecting patterns across areas, allowing us to train models that can predict and suggest crime prevention steps to individual businesses even before they start their store, personalised based on location, type of store and images.