Developed during the Fall 2022 Waterloo.AI Data Challenge sponsored by Musashi AI, this project focuses on detecting human presence in fisheye top-down surveillance footage for workplace monitoring. The goal of the project is to accurately identify humans within predefined regions of interest (ROIs) using ensemble machine learning techniques tailored for complex environments captured through fisheye cameras.

Our solution uses a combination of pre-trained Convolutional Neural Networks (CNNs), specifically ResNet-101 for robust feature extraction, and unsupervised clustering methods like k-Means to segment and detect human figures. This ensemble approach achieves efficient detection and monitoring in dynamic workplace environments, leading to both accuracy and computational efficiency.

Recognition & Outreach
• 2nd Place Winner in Waterloo.AI Data Challenge (1,000 CAD Award)
• UWaterloo SCS News: Cheriton School of Computer Science undergrads and grads among winners at Waterloo.AI’s Data Challenge
• Waterloo.AI News: Link

Core Features
• Utilizes pre-trained ResNet-101 to extract features from image data.
• Applies k-Means clustering to group data points and identify human figures in the ROIs.
• Explores multiple approaches including CNN fine-tuning and linear classifiers to improve detection accuracy.

• SkillsResNet-101, k-Means Clustering, CNN Fine-tuning, Image Processing, Model Training
• Team MembersYuzhe You, Mohammad Zarei, Ce Zhang
• KeywordsComputer Vision, Human Detection, ResNet-101, Clustering, Image Processing, ROI Detection