Loading page
Academic & AI researchAI & ML
Object & Face-Detection System
Real-time computer vision boundary mapping and multi-class object localization.
Headline results
- FPS
- 30+ FPS
- Precision
- 91.2% mAP
The project in 7 images
5 summary slides · 2 genuine images from the project
01Overview
The project at a glance
A lightweight AI system designed for security cameras, autonomous systems, and edge nodes, locating and identifying human faces and multiple object classes.
- Type
- Academic & AI research
- Category
- AI & ML
Tech stack
- Python
- TensorFlow
- OpenCV
- NumPy
- Deep Learning
02The challenge
The research question
Standard CNN models are computationally heavy, requiring dedicated GPUs that are unavailable on IoT sensors and basic security camera feeds.
The brief
Running object classification models locally on low-power devices and edge cameras, requiring low latency and low power consumption.
03The solution
The model pipeline
Developed an optimized computer vision execution pipeline using OpenCV and pre-trained deep neural network architectures. Trimmed convolutional layers and set up multithreaded frame rendering queues.
Loads model configurations via OpenCV's DNN module. Frames are captured, downsampled, processed via MobileNet-SSD parameters, and overlaid with labels in real-time.
- 01
Data
- OpenCV
- NumPy
- 02
Model
- TensorFlow
- 03
Evaluation
- 30+ FPS
- Precision 91.2% mAP
- 04
Output
- AI & ML
04Key features
What it does
- 01Real-time bounding box tracking for multi-class targets.
- 02Face boundary mapping resilient to angle shifts.
- 03Zero network-dependency processing (fully local deployment).
05Results
What it achieved
Achieved steady 30+ FPS real-time detection on standard CPU hardware with an average Mean Average Precision (mAP) of 91.2%.
Next step
Want AI like this in your product?
From data and model training to evaluation and a production API. Tell me what you want it to do.
Back to all work