real-time-det.py

# /// script
# dependencies = [
#   "transformers",
#   "torch",
#   "opencv-python",
#   "Pillow",
# ]
# ///

import time
import cv2
import torch
from transformers import AutoImageProcessor, RTDetrForObjectDetection

device = "mps"
model_id = "PekingU/rtdetr_r50vd_coco_o365"
threshold = 0.4
scale = 0.5

processor = AutoImageProcessor.from_pretrained(model_id)
model = RTDetrForObjectDetection.from_pretrained(model_id).to(device)
model.eval()

def draw_boxes(frame, boxes):
    for (x1, y1, x2, y2) in boxes:
        cv2.rectangle(frame, (int(x1), int(y1)), (int(x2), int(y2)), (0, 255, 255), 2)

cap = cv2.VideoCapture(0)

while True:
    start_time = time.time()

    ret, frame = cap.read()
    if not ret:
        break

    frame = cv2.resize(frame, None, fx=scale, fy=scale)
    rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)

    inputs = processor(images=rgb, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = model(**inputs)

    h, w = frame.shape[:2]
    results = processor.post_process_object_detection(
        outputs,
        target_sizes=torch.tensor([(h, w)]),
        threshold=threshold,
    )[0]

    person_boxes = results["boxes"][results["labels"] == 0].cpu().numpy()
    draw_boxes(frame, person_boxes)

    fps = 1.0 / (time.time() - start_time)
    cv2.putText(frame, f"FPS: {fps:.1f}", (10, 30),
                cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 255, 0), 2)

    cv2.imshow("Person Detection", frame)
    if cv2.waitKey(1) & 0xFF in [27, ord("q")]:
        break

cap.release()
cv2.destroyAllWindows()