kovelja009/hand_wave_detection

Hand wave detector which uses custom yolov5s as a backbone for hand detection and random forest for wave classifier.

Table of Contents

• Hand Detection
• Hand Tracking
• Data Collection
• Feature Engineering
• Classification

---

Hand Detection

Dataset	Number of images	Number of classes	Classes	Epochs
HaGRID micro 3 classes	1.3k	3	Fist, Palm, No Gesture	338/1000 (early stopping after 100)

---

Hand tracking

• SORT (Simple Online and Realtime Tracking)
• Multiple objects tracker (waits for 20 frames before terminating track)

---

Data collection

Class	Training	Test
wave	video 1
wave	video 2
wave	video 3
non-wave	video 4
non-wave	video 5
wave		video 6
non-wave		video 7

• Training set: 287 wave and 321 non-wave signals
• Test set: 144 wave and 335 non-wave signals

---

Feature engineering

• Features of the signal (x - blue, y - green)
• Signal is of the length of 40 frames (camera runs on about 30 fps)

---

Classification

• Random forest classificator
• 13 trees
• Test accuracy: 0.93