Our purpose is to transform public safety and provide justice by using AI powered tools, ensuring that evidence is easily accessible to authorities. We believe that justice is not being properly provided to all communities mainly due to higher and lower crime rates as well as systemic bias. CI4 serves as an unbiased and overall reliable witness that protects the innocent, moving us towards a justice system that promotes equality.

We created this tool using a variety of tools.

We combined state-of-the-art AI methods with specific datasets to create a tool that is both accurate and fair. At the heart of CI4 is the X3D model, a cutting-edge machine learning architecture designed for video understanding.

We custom-trained this model using two complementary datasets:

UCF Anomaly Detection Dataset – containing real crime-related video scenarios, which provided the ground truth for crime recognition.

UCF101 Dataset – a large-scale action recognition dataset used as a control set, ensuring the system could also learn what normal, non-criminal activities look like.

Our training pipeline was designed as a multi-task system with two specialized “heads”:

Binary Classifier Head: responsible for determining whether a crime occurred within an entire video. This gave us a reliable yes/no answer to crime detection.

Crime Classification Head: a more advanced head that operated on short clips of 16 frames each. By analyzing these clips, the model learned to identify the specific type of crime (e.g., robbery, assault, burglary).

To enhance accuracy, we introduced context windows powered by transformers. These allowed the model to understand the temporal flow of events: if one clip was labeled as part of a robbery, the surrounding clips were also likely to belong to that event. This context-aware learning significantly improved our ability to pinpoint the exact start and end of a crime, rather than treating a video as a single undivided unit.

Together, these innovations allowed CI4 to function as an AI-powered unbiased witness

Because X3D is computationally efficient, CI4 could operate in real-time on surveillance camera feeds, providing authorities with immediate alerts and accessible evidence.

Built With

Share this project:

Updates