Project Overview
Collaborated with the Signal Hill Police Department to develop AI-driven software for identifying and categorizing police body-worn camera footage eligible for purging.
AI-powered classification system for police body camera footage
Tech Stack
Key Features
Transcript Generator
Implemented a transcript generator using faster-whisper model, adding compression, volume, and speech-pattern thresholds for higher accuracy.
Zero-shot LLM Classifier
Uses transcribed audio to classify video category, leveraging large language models to understand context without specific training.
Video Similarity Model
S3D Video Model trained on similar videos within each category and produces a similarity score for each video, enabling accurate categorization.
Audio ML Model
Building an RCNN that detects common audio events such as police sirens, enhancing classification accuracy through sound recognition.
Object Detection Model
Detects objects within a video such as police car doors and equipment, providing visual cues for more accurate video classification.
Applications
- Identify and tag a backlog of untagged police videos
- Use as an AI suggestion for manual police tagging software
- Generalize video categorization techniques to any video classification, filter, or search software