Task 6

Object Detection Model Training

ai-model/

A YOLOv8n model trained from scratch — not a pretrained checkpoint — on a self-assembled ~30,000-image dataset covering fire, person, and smoke detection.

What was asked

Collect a sufficient number of images for training and testing an object-detection model using TensorFlow, OpenCV, or YOLO. Training your own model is required — using a pre-existing model is explicitly not sufficient.

Dataset

Sourced and assembled via Roboflow: fire-and-person-surveillance-detection (version 2), 3 classes.

Split	Images	fire	person	smoke
train	21,831	4,773	73,646	336
valid	5,488	663	20,966	173
test	2,441	237	10,840	37
total	29,760	5,673	105,452	546

The class distribution is heavily imbalanced — person instances outnumber smoke roughly 193:1 — which is a plausible explanation for smoke's lower recall (see below).

Training pipeline

Parameter	Default	Description
`EPOCHS`	300	maximum training epochs
`IMG_SIZE`	640	input image size
`BATCH`	16	batch size
`PATIENCE`	50	early-stopping patience
`LR0` / `LRF`	0.01 / 0.1	initial / final learning rate

train/train.py downloads the dataset, trains YOLOv8n with early stopping, and exports the best weights to ONNX. detect/detect_webcam.py runs live webcam inference with an FPS overlay for local testing before deployment to the Hailo accelerator (see the Inference node). Device selection auto-detects CUDA → Apple MPS → CPU.

Results

Precision

0.843

Recall

0.695

mAP@50

0.772

mAP@50-95

0.493

Precision (84%) — of all boxes the model predicts, 84% are correct.
Recall (70%) — of all actual fire/person instances present, the model catches about 70%; roughly 30% are missed.
mAP@50 (77%) — mean average precision at a loose 0.5 IoU threshold: a detection only needs to roughly cover the object to count as correct.
mAP@50-95 (49%) — the same metric at much stricter overlap thresholds. The gap versus mAP@50 shows the model reliably finds objects but doesn't always draw tight boxes — typical for the nano-sized YOLOv8 variant chosen for edge inference.

Per-class breakdown

Class	Images	Instances	Precision	Recall	mAP@50	mAP@50-95
all	5,488	27,289	0.842	0.697	0.772	0.493
fire	1,060	1,667	0.799	0.704	0.757	0.428
person	3,834	24,799	0.902	0.755	0.843	0.604
smoke	715	823	0.824	0.633	0.718	0.447

smoke has the lowest recall (63%) and mAP@50-95 (0.447) of the three classes — consistent with it being by far the most underrepresented in training (546 instances vs. 105k for person). fire performs closer to the class average despite also being a minority class.

Deployment target

The trained model is exported to ONNX and compiled for the Hailo-8L accelerator that powers the inference node — a Pi 5 + Hailo AI HAT+ pairing that runs real-time inference on the sensor node's MJPEG stream and reports detections to the backend (Task 7).