SportsAI / src /exercises /dead_lift.py
nicolasbuitragob's picture
dead lift exercise
478a673
raw
history blame
11.5 kB
import cv2
import numpy as np
from scipy.optimize import linear_sum_assignment
import requests
import json
from fastapi.responses import JSONResponse
from config import API_URL, API_KEY
import logging
logger = logging.getLogger(__name__)
class RepCounter:
def __init__(self, fps, height_cm, mass_kg=0, target_reps=None):
self.count = 0
self.last_state = None
self.cooldown_frames = 15
self.cooldown = 0
self.rep_start_frame = None
self.start_wrist_y = None
self.rep_data = []
self.power_data = []
self.fps = fps
self.cm_per_pixel = None
self.real_distance_cm = height_cm * 0.2735
self.calibration_done = False
self.mass_kg = mass_kg
self.gravity = 9.81
self.target_reps = int(target_reps)
self.target_reached = False
self.final_speed = None
self.final_power = None
self.SKELETON = [
(5, 6), (5, 7), (7, 9), (6, 8), (8, 10),
(9, 10), (11, 12), (5, 11), (6, 12),
(11, 13), (13, 15), (12, 14), (14, 16), (13, 14)
]
def update(self, wrist_y, knee_y, current_frame):
if self.target_reached or self.cooldown > 0:
self.cooldown = max(0, self.cooldown - 1)
return
current_state = 'above' if wrist_y < knee_y else 'below'
if self.last_state != current_state:
if current_state == 'below':
self.rep_start_frame = current_frame
self.start_wrist_y = wrist_y
elif current_state == 'above' and self.last_state == 'below':
if self.rep_start_frame is not None and self.cm_per_pixel is not None:
end_frame = current_frame
duration = (end_frame - self.rep_start_frame) / self.fps
distance_pixels = self.start_wrist_y - wrist_y
distance_cm = distance_pixels * self.cm_per_pixel
if duration > 0:
speed_cmps = abs(distance_cm) / duration
self.rep_data.append(speed_cmps)
if self.mass_kg > 0:
speed_mps = speed_cmps / 100
force = self.mass_kg * self.gravity
power = force * speed_mps
self.power_data.append(power)
self.count += 1
if self.target_reps and self.count >= self.target_reps:
self.count = self.target_reps
self.target_reached = True
self.final_speed = np.mean(self.rep_data) if self.rep_data else 0
self.final_power = np.mean(self.power_data) if self.power_data else 0
self.cooldown = self.cooldown_frames
self.last_state = current_state
# CentroidTracker class
class CentroidTracker:
def __init__(self, max_disappeared=50, max_distance=100):
self.next_id = 0
self.objects = {}
self.max_disappeared = max_disappeared
self.max_distance = max_distance
def _update_missing(self):
to_delete = []
for obj_id in list(self.objects.keys()):
self.objects[obj_id]["missed"] += 1
if self.objects[obj_id]["missed"] > self.max_disappeared:
to_delete.append(obj_id)
for obj_id in to_delete:
del self.objects[obj_id]
def update(self, detections):
if len(detections) == 0:
self._update_missing()
return []
centroids = np.array([[(x1 + x2) / 2, (y1 + y2) / 2] for x1, y1, x2, y2 in detections])
if len(self.objects) == 0:
return self._register_new(centroids)
return self._match_existing(centroids, detections)
def _register_new(self, centroids):
new_ids = []
for centroid in centroids:
self.objects[self.next_id] = {"centroid": centroid, "missed": 0}
new_ids.append(self.next_id)
self.next_id += 1
return new_ids
def _match_existing(self, centroids, detections):
existing_ids = list(self.objects.keys())
existing_centroids = [self.objects[obj_id]["centroid"] for obj_id in existing_ids]
cost = np.linalg.norm(np.array(existing_centroids)[:, np.newaxis] - centroids, axis=2)
row_ind, col_ind = linear_sum_assignment(cost)
used_rows = set()
used_cols = set()
matches = {}
for (row, col) in zip(row_ind, col_ind):
if cost[row, col] <= self.max_distance:
obj_id = existing_ids[row]
matches[obj_id] = centroids[col]
used_rows.add(row)
used_cols.add(col)
for obj_id in existing_ids:
if obj_id not in matches:
self.objects[obj_id]["missed"] += 1
if self.objects[obj_id]["missed"] > self.max_disappeared:
del self.objects[obj_id]
new_ids = []
for col in range(len(centroids)):
if col not in used_cols:
self.objects[self.next_id] = {"centroid": centroids[col], "missed": 0}
new_ids.append(self.next_id)
self.next_id += 1
for obj_id, centroid in matches.items():
self.objects[obj_id]["centroid"] = centroid
self.objects[obj_id]["missed"] = 0
all_ids = []
for detection in detections:
centroid = np.array([(detection[0] + detection[2]) / 2, (detection[1] + detection[3]) / 2])
min_id = None
min_dist = float('inf')
for obj_id, data in self.objects.items():
dist = np.linalg.norm(centroid - data["centroid"])
if dist < min_dist and dist <= self.max_distance:
min_dist = dist
min_id = obj_id
if min_id is not None:
all_ids.append(min_id)
self.objects[min_id]["centroid"] = centroid
else:
all_ids.append(self.next_id)
self.objects[self.next_id] = {"centroid": centroid, "missed": 0}
self.next_id += 1
return all_ids
# Función de procesamiento optimizada
def process_frame_for_counting(frame, tracker, rep_counter, frame_number,vitpose):
pose_results = vitpose.pipeline(frame)
keypoints = pose_results.keypoints_xy.float().cpu().numpy()[0]
scores = pose_results.scores.float().cpu().numpy()[0]
valid_points = {}
wrist_midpoint = None
knee_line_y = None
print(keypoints)
print(scores)
# Procesar puntos clave
for i, (kp, conf) in enumerate(zip(keypoints, scores)):
if conf > 0.3 and 5 <= i <= 16:
x, y = map(int, kp[:2])
valid_points[i] = (x, y)
# Calibración usando keypoints de rodilla (14) y pie (16)
if not rep_counter.calibration_done and 14 in valid_points and 16 in valid_points:
knee = valid_points[14]
ankle = valid_points[16]
pixel_distance = np.sqrt((knee[0] - ankle[0])**2 + (knee[1] - ankle[1])**2)
if pixel_distance > 0:
rep_counter.cm_per_pixel = rep_counter.real_distance_cm / pixel_distance
rep_counter.calibration_done = True
# Calcular puntos de referencia para conteo
if 9 in valid_points and 10 in valid_points:
wrist_midpoint = (
(valid_points[9][0] + valid_points[10][0]) // 2,
(valid_points[9][1] + valid_points[10][1]) // 2
)
if 13 in valid_points and 14 in valid_points:
pt1 = np.array(valid_points[13])
pt2 = np.array(valid_points[14])
direction = pt2 - pt1
extension = 0.2
new_pt1 = pt1 - direction * extension
new_pt2 = pt2 + direction * extension
knee_line_y = (new_pt1[1] + new_pt2[1]) // 2
# Actualizar contador
if wrist_midpoint and knee_line_y:
rep_counter.update(wrist_midpoint[1], knee_line_y, frame_number)
# Función principal de Gradio
def analyze_dead_lift(input_video, reps, weight, height,vitpose,player_id,exercise_id):
cap = cv2.VideoCapture(input_video)
fps = cap.get(cv2.CAP_PROP_FPS)
rep_counter = RepCounter(fps, int(height), int(weight), int(reps))
tracker = CentroidTracker(max_distance=150)
frame_number = 0
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
process_frame_for_counting(frame, tracker, rep_counter, frame_number,vitpose)
frame_number += 1
cap.release()
# Preparar payload para webhook
if rep_counter.mass_kg > 0:
power_data = rep_counter.power_data
else:
# Si no hay masa, usar ceros para potencia
power_data = [0] * len(rep_counter.rep_data) if rep_counter.rep_data else []
# Asegurar que tenemos datos para enviar
if rep_counter.rep_data:
payload = {"repetition_data": [
{"repetition": i, "velocidad": round(s,1), "potencia": round(p,1)}
for i, (s, p) in enumerate(zip(rep_counter.rep_data, power_data), start=1)
]}
else:
# En caso de no detectar repeticiones
payload = {"repetition_data": []}
send_results_api(payload, player_id, exercise_id, input_video)
def send_results_api(results_dict: dict,
player_id: str,
exercise_id: str,
video_path: str) -> JSONResponse:
"""
Send video analysis results to the API webhook endpoint.
This function uploads the analyzed video file along with the computed metrics
to the API's webhook endpoint for processing and storage.
Args:
results_dict (dict): Dictionary containing analysis results including:
- video_analysis: Information about the processed video
- repetition_data: List of metrics for each jump repetition
player_id (str): Unique identifier for the player
exercise_id (str): Unique identifier for the exercise
video_path (str): Path to the video file to upload
Returns:
JSONResponse: HTTP response from the API endpoint
Raises:
FileNotFoundError: If the video file doesn't exist
requests.RequestException: If the API request fails
json.JSONEncodeError: If results_dict cannot be serialized to JSON
"""
url = API_URL + "/exercises/webhooks/video-processed-results"
logger.info(f"Sending video results to {url}")
# Open the video file
with open(video_path, 'rb') as video_file:
# Prepare the files dictionary for file upload
files = {
'file': (video_path.split('/')[-1], video_file, 'video/mp4')
}
# Prepare the form data
data = {
'player_id': player_id,
'exercise_id': exercise_id,
'results': json.dumps(results_dict) # Convert dict to JSON string
}
# Send the request with both files and data
response = requests.post(
url,
headers={"token": API_KEY},
files=files,
data=data,
stream=True
)
logger.info(f"Response: {response.status_code}")
logger.info(f"Response: {response.text}")
return response