Real-time Object Detection System
Real-time object recognition from a webcam feed
- YOLOv3
- Model
- Gerçek zamanlı
- Processing
- Açık kaynak
- License
Overview
The real difficulty in object detection is not accuracy but holding that accuracy in real time. Labelling a single photograph correctly and tracking thirty frames a second are very different engineering problems.
This application was built on pre-trained YOLOv3 weights. OpenCV pulls the camera stream frame by frame, each frame is passed to the network, and the returned bounding boxes are filtered by confidence score. Overlapping boxes produced for the same object are reduced to one with non-maximum suppression; without that step the screen fills with dozens of boxes for a single object.
For performance, the input resolution and confidence threshold were left adjustable, so the system can be balanced for cases that need high accuracy or a smooth frame rate. The code is shared open source on GitHub.
Highlights
- Frame-by-frame real-time detection on a camera stream
- Overlapping boxes cleaned up with non-maximum suppression
- Adjustable confidence threshold and input resolution
- Shared open source on GitHub
Objectives
- Adapt a pre-trained model to a real-time stream
- Create an adjustable balance between accuracy and frame rate
- Run on an ordinary computer
Solution
- Loading YOLOv3 weights through the OpenCV DNN module
- Confidence filtering and NMS implementation
- Optimising the frame processing loop
- Exposing parameters for external tuning
Approach
- 01Evaluating models and weights
- 02Performance testing on a live stream
- 03Preparing the code for open source release
Outcomes
- Smooth real-time detection achieved with a standard webcam
- Adaptable to different hardware through its parameters
- The project became a reference example for computer vision learners
Challenges
- Maintaining frame rate under real-time processing
- Cleaning up overlapping detection boxes
- Detection stability in low light
Selected Work
Similar projects
Rice Classification Model
A deep learning model trained with TensorFlow and Keras that classifies rice varieties from grain images, reaching 98% accuracy on the test set.
Audio-to-News-Article Agent
An AI tool that turns an audio recording first into text, then into a structured news article, combining speech recognition and natural language processing in a single pipeline.
PDF Q&A Agent (ChatPDF)
A RAG-based assistant that analyses uploaded PDFs and answers user questions grounded in the document's own content.