View source1 Background and motivation
This Urban Futurism competition project explored whether a camera system could detect useful activity while limiting how much identifying imagery appeared in the displayed feed. The objective joined two concerns that are often treated separately: recognizing what is happening in a scene and deciding which visual details the viewer actually needs to see.
2 Vision pipeline
I built the core vision pipeline using YOLO for object detection and OpenCV for tracking and region logic. Detection identifies objects within a frame, while tracking and regions provide context over time and space. This allows the application to reason about activity in a selected area instead of treating every frame as an unrelated image.
The implementation had to connect capture, processing, and display in a live loop. Work that is acceptable in an offline notebook can create delay when applied to every frame, so the practical task was to keep the scene interpretable while processing continued. The portfolio’s project record describes this as a real-time prototype rather than supplying a hardware-independent frame-rate benchmark.
3 Privacy filtering
We experimented with masking and blur treatments, including generative approaches, to obscure sensitive parts of the feed while retaining useful scene context. The design question was how much visual information could be removed without making the activity view unusable. That creates a dependency between detection and display: a filter can only affect the regions that the pipeline identifies and processes.
Masking is therefore a presentation technique, not a complete guarantee of anonymity. Missed detections, recognizable clothing, or surrounding context can still identify someone. The project explored the rendering boundary, and its title should be understood as a design aim rather than a claim that the complete system had been privacy-audited.
4 Outcome
The team ranked seventh in the competition. For me, the main result was experience connecting vision models to an interactive system under live processing constraints. The project brought detection, temporal handling, region logic, and visual filtering into the same loop, making their interactions visible beyond the accuracy of an individual model.