I worked with the CAOE to improve an in-development computer vision system to record the flow of crowds with CCTV cameras. I led a team of student researchers who identified weaknesses in the existing solution which scored 42% accurate against benchmarks. We evaluated the implementation, identified key weaknesses and designed a new system. Delivered system produced data that was 94% accurate compared to the hand-labeled test dataset (resolution = 1 minute).

Key Deliverables:

• Unique subject identification

• Short-term memory system

• Subject occlusion resolution

• Dynamic mapping of key areas for use in region-based custom metrics capture

• Creation of a hand-labeled temporal evaluation dataset for foot traffic spanning 72 hours of footage

• Benchmark and evaluation suite