Real-Time Object Detection

Problem

Academic exploration of real-time computer vision: could a lightweight detection model identify objects and people in a live video stream fast enough to be practically useful, without requiring a GPU-only production pipeline?

Solution

A real-time detection pipeline using YOLOv4-Tiny on the Darknet framework, configured against the COCO dataset, drawing bounding boxes with class labels and confidence scores over a live video feed via OpenCV.

Architecture

Darknet loads the YOLOv4-Tiny configuration and pre-trained weights; OpenCV handles video capture, frame preprocessing, and rendering of detection overlays (bounding boxes, class names, confidence scores) back onto each frame.

Challenges

• —Balancing inference speed against detection accuracy by choosing the -Tiny model variant
• —Tuning confidence thresholds to reduce false positives on a live camera feed
• —Working within Darknet's configuration-driven workflow rather than a Python-native training loop

Learnings

• —Hands-on understanding of one-stage detector trade-offs (speed vs. accuracy)
• —How model-size choice (Tiny vs. full YOLOv4) directly impacts real-time feasibility
• —Practical experience with the classic Darknet/OpenCV computer-vision stack