Update app.py
Browse files
app.py
CHANGED
@@ -1,287 +1,133 @@
|
|
1 |
import cv2
|
2 |
import numpy as np
|
3 |
-
import pandas as pd
|
4 |
-
import plotly.express as px
|
5 |
-
import plotly.graph_objects as go
|
6 |
import torch
|
|
|
7 |
import gradio as gr
|
|
|
|
|
8 |
import os
|
9 |
-
import time
|
10 |
-
from scipy.optimize import curve_fit
|
11 |
-
import sys
|
12 |
|
13 |
-
#
|
14 |
-
|
15 |
|
16 |
-
#
|
17 |
-
|
18 |
-
|
|
|
19 |
|
20 |
-
# Cricket pitch dimensions (in meters)
|
21 |
-
PITCH_LENGTH = 20.12 # Length of cricket pitch (stumps to stumps)
|
22 |
-
PITCH_WIDTH = 3.05 # Width of pitch
|
23 |
-
STUMP_HEIGHT = 0.71 # Stump height
|
24 |
-
STUMP_WIDTH = 0.2286 # Stump width (including bails)
|
25 |
-
|
26 |
-
# Model input size (adjust if yolov5s.pt was trained with a different size)
|
27 |
-
MODEL_INPUT_SIZE = (640, 640) # (height, width)
|
28 |
-
FRAME_SKIP = 2 # Process every 2nd frame
|
29 |
-
MIN_DETECTIONS = 10 # Stop after 10 detections
|
30 |
-
BATCH_SIZE = 4 # Process 4 frames at a time
|
31 |
-
SLOW_MOTION_FACTOR = 3 # Duplicate each frame 3 times for slow motion
|
32 |
-
|
33 |
-
# Load model
|
34 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
35 |
-
model = attempt_load("best.pt") # Load yolov5s.pt
|
36 |
-
model.to(device).eval() # Move model to device and set to evaluation mode
|
37 |
-
|
38 |
-
# Function to process video and detect ball
|
39 |
def process_video(video_path):
|
|
|
40 |
cap = cv2.VideoCapture(video_path)
|
41 |
-
|
42 |
-
|
43 |
-
frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
44 |
-
positions = []
|
45 |
-
frame_numbers = []
|
46 |
-
bounce_frame = None
|
47 |
-
bounce_point = None
|
48 |
-
batch_frames = []
|
49 |
-
batch_frame_nums = []
|
50 |
-
frame_count = 0
|
51 |
|
52 |
-
start_time = time.time()
|
53 |
while cap.isOpened():
|
54 |
-
frame_num = int(cap.get(cv2.CAP_PROP_POS_FRAMES))
|
55 |
ret, frame = cap.read()
|
56 |
if not ret:
|
57 |
break
|
58 |
-
|
59 |
-
#
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
frame_count += 1
|
69 |
-
|
70 |
-
# Process batch when full or at end
|
71 |
-
if len(batch_frames) == BATCH_SIZE or not ret:
|
72 |
-
# Preprocess batch
|
73 |
-
batch = [cv2.cvtColor(f, cv2.COLOR_BGR2RGB) for f in batch_frames]
|
74 |
-
batch = np.stack(batch) # [batch_size, H, W, 3]
|
75 |
-
batch = torch.from_numpy(batch).to(device).float() / 255.0
|
76 |
-
batch = batch.permute(0, 3, 1, 2) # [batch_size, 3, H, W]
|
77 |
-
|
78 |
-
# Run inference
|
79 |
-
frame_start_time = time.time()
|
80 |
-
with torch.no_grad():
|
81 |
-
pred = model(batch)[0]
|
82 |
-
pred = non_max_suppression(pred, conf_thres=0.25, iou_thres=0.45)
|
83 |
-
print(f"Batch inference time: {time.time() - frame_start_time:.2f}s for {len(batch_frames)} frames")
|
84 |
-
|
85 |
-
# Process detections
|
86 |
-
for i, det in enumerate(pred):
|
87 |
-
if det is not None and len(det):
|
88 |
-
det = xywh2xyxy(det) # Convert to [x1, y1, x2, y2]
|
89 |
-
for *xyxy, conf, cls in det:
|
90 |
-
x_center = (xyxy[0] + xyxy[2]) / 2
|
91 |
-
y_center = (xyxy[1] + xyxy[3]) / 2
|
92 |
-
# Scale coordinates back to original frame size
|
93 |
-
x_center = x_center * frame_width / MODEL_INPUT_SIZE[1]
|
94 |
-
y_center = y_center * frame_height / MODEL_INPUT_SIZE[0]
|
95 |
-
positions.append((x_center.item(), y_center.item()))
|
96 |
-
frame_numbers.append(batch_frame_nums[i])
|
97 |
-
|
98 |
-
# Detect bounce (lowest y_center point)
|
99 |
-
if bounce_frame is None or y_center > positions[bounce_frame][1]:
|
100 |
-
bounce_frame = len(frame_numbers) - 1
|
101 |
-
bounce_point = (x_center.item(), y_center.item())
|
102 |
-
|
103 |
-
batch_frames = []
|
104 |
-
batch_frame_nums = []
|
105 |
-
|
106 |
-
# Early termination
|
107 |
-
if len(positions) >= MIN_DETECTIONS:
|
108 |
-
break
|
109 |
-
|
110 |
cap.release()
|
111 |
-
print(f"Total video processing time: {time.time() - start_time:.2f}s")
|
112 |
-
return positions, frame_numbers, bounce_point, frame_rate, frame_width, frame_height
|
113 |
-
|
114 |
-
# Polynomial function for trajectory fitting
|
115 |
-
def poly_func(x, a, b, c):
|
116 |
-
return a * x**2 + b * x + c
|
117 |
|
118 |
-
|
119 |
-
def predict_trajectory(positions, frame_numbers, frame_width, frame_height):
|
120 |
-
if len(positions) < 3:
|
121 |
-
return None, None, "Insufficient detections for trajectory prediction"
|
122 |
|
123 |
-
|
124 |
-
|
125 |
-
|
|
|
|
|
|
|
|
|
|
|
126 |
|
127 |
-
#
|
128 |
try:
|
129 |
-
|
130 |
-
|
131 |
except:
|
132 |
-
return None, None
|
133 |
-
|
134 |
-
#
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
#
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
distances = []
|
172 |
-
for i in range(1, len(positions)):
|
173 |
-
x1, y1 = positions[i-1]
|
174 |
-
x2, y2 = positions[i]
|
175 |
-
pixel_dist = np.sqrt((x2 - x1)**2 + (y2 - y1)**2)
|
176 |
-
distances.append(pixel_dist)
|
177 |
-
|
178 |
-
pixel_to_meter = PITCH_LENGTH / frame_width
|
179 |
-
distances_m = [d * pixel_to_meter for d in distances]
|
180 |
-
time_interval = 1 / frame_rate
|
181 |
-
speeds = [d / time_interval for d in distances_m]
|
182 |
-
avg_speed_kmh = np.mean(speeds) * 3.6
|
183 |
-
return avg_speed_kmh, "Speed calculated successfully"
|
184 |
-
|
185 |
-
# Main Gradio function with video overlay and slow motion
|
186 |
-
def drs_analysis(video):
|
187 |
-
# Video is a file path (string) in Hugging Face Spaces
|
188 |
-
video_path = video if isinstance(video, str) else "temp_video.mp4"
|
189 |
-
if not isinstance(video, str):
|
190 |
-
with open(video_path, "wb") as f:
|
191 |
-
f.write(video.read())
|
192 |
-
|
193 |
-
# Process video for detections
|
194 |
-
positions, frame_numbers, bounce_point, frame_rate, frame_width, frame_height = process_video(video_path)
|
195 |
-
if not positions:
|
196 |
-
return None, None, "No ball detected in video", None
|
197 |
-
|
198 |
-
# Predict trajectory and wicket path
|
199 |
-
trajectory, inline_path, lbw_decision = predict_trajectory(positions, frame_numbers, frame_width, frame_height)
|
200 |
-
if trajectory is None:
|
201 |
-
return None, None, lbw_decision, None
|
202 |
-
|
203 |
-
pitch_x, pitch_y = map_pitch(bounce_point, frame_width, frame_height)
|
204 |
-
speed_kmh, speed_status = estimate_speed(positions, frame_numbers, frame_rate, frame_width)
|
205 |
-
|
206 |
-
# Create output video with overlays and slow motion
|
207 |
-
output_path = "output_video.mp4"
|
208 |
-
cap = cv2.VideoCapture(video_path)
|
209 |
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
210 |
-
out = cv2.VideoWriter(output_path, fourcc,
|
211 |
-
|
212 |
-
frame_count = 0
|
213 |
-
positions_dict = dict(zip(frame_numbers, positions))
|
214 |
|
215 |
-
|
216 |
-
ret, frame = cap.read()
|
217 |
-
if not ret:
|
218 |
-
break
|
219 |
-
|
220 |
-
# Skip frames for consistency with detection
|
221 |
-
if frame_count % FRAME_SKIP != 0:
|
222 |
-
frame_count += 1
|
223 |
-
continue
|
224 |
-
|
225 |
-
# Overlay ball trajectory (red) and wicket inline path (blue)
|
226 |
-
if frame_count in positions_dict:
|
227 |
-
cv2.circle(frame, (int(positions_dict[frame_count][0]), int(positions_dict[frame_count][1])), 5, (0, 0, 255), -1) # Red dot
|
228 |
if trajectory:
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
cv2.line(frame, (traj_x[i-1], traj_y[i-1]), (traj_x[i], traj_y[i]), (0, 0, 255), 2) # Red line
|
234 |
-
if inline_path:
|
235 |
-
inline_x = [int(x) for x, _ in inline_path]
|
236 |
-
inline_y = [int(y) for _, y in inline_path]
|
237 |
-
if inline_x and inline_y:
|
238 |
-
for i in range(1, len(inline_x)):
|
239 |
-
cv2.line(frame, (inline_x[i-1], inline_y[i-1]), (inline_x[i], inline_y[i]), (255, 0, 0), 2) # Blue line
|
240 |
-
|
241 |
-
# Overlay pitch map in top-right corner
|
242 |
-
if pitch_x is not None and pitch_y is not None:
|
243 |
-
map_width = 200
|
244 |
-
# Cap map_height to 25% of frame height to ensure it fits
|
245 |
-
map_height = min(int(map_width * PITCH_LENGTH / PITCH_WIDTH), frame_height // 4)
|
246 |
-
pitch_map = np.zeros((map_height, map_width, 3), dtype=np.uint8)
|
247 |
-
pitch_map[:] = (0, 255, 0) # Green pitch
|
248 |
-
cv2.rectangle(pitch_map, (0, map_height-10), (map_width, map_height), (0, 51, 51), -1) # Brown stumps
|
249 |
-
bounce_x = int((pitch_x + PITCH_WIDTH/2) / PITCH_WIDTH * map_width)
|
250 |
-
bounce_y = int((1 - pitch_y / PITCH_LENGTH) * map_height)
|
251 |
-
cv2.circle(pitch_map, (bounce_x, bounce_y), 5, (0, 0, 255), -1) # Red bounce point
|
252 |
-
# Ensure overlay fits within frame
|
253 |
-
overlay_region = frame[0:map_height, frame_width-map_width:frame_width]
|
254 |
-
if overlay_region.shape[0] >= map_height and overlay_region.shape[1] >= map_width:
|
255 |
-
frame[0:map_height, frame_width-map_width:frame_width] = cv2.resize(pitch_map, (map_width, map_height))
|
256 |
-
|
257 |
-
# Add text annotations
|
258 |
-
text = f"LBW: {lbw_decision}\nSpeed: {speed_kmh:.2f} km/h"
|
259 |
-
cv2.putText(frame, text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2, cv2.LINE_AA)
|
260 |
-
|
261 |
-
# Write frame multiple times for slow motion
|
262 |
-
for _ in range(SLOW_MOTION_FACTOR):
|
263 |
-
out.write(frame)
|
264 |
-
|
265 |
-
frame_count += 1
|
266 |
-
|
267 |
-
cap.release()
|
268 |
out.release()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
269 |
|
270 |
-
|
271 |
-
|
|
|
272 |
|
273 |
-
return
|
274 |
|
275 |
# Gradio interface
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
|
|
285 |
|
286 |
if __name__ == "__main__":
|
287 |
-
|
|
|
1 |
import cv2
|
2 |
import numpy as np
|
|
|
|
|
|
|
3 |
import torch
|
4 |
+
from ultralytics import YOLO
|
5 |
import gradio as gr
|
6 |
+
from scipy.interpolate import interp1d
|
7 |
+
import uuid
|
8 |
import os
|
|
|
|
|
|
|
9 |
|
10 |
+
# Load the trained YOLOv8n model from the Space's root directory
|
11 |
+
model = YOLO("best.pt") # Assumes best.pt is in the same directory as app.py
|
12 |
|
13 |
+
# Constants for LBW decision
|
14 |
+
STUMPS_WIDTH = 0.2286 # meters (width of stumps)
|
15 |
+
BALL_DIAMETER = 0.073 # meters (approx. cricket ball diameter)
|
16 |
+
FRAME_RATE = 30 # Default frame rate for video processing
|
17 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
def process_video(video_path):
|
19 |
+
# Initialize video capture
|
20 |
cap = cv2.VideoCapture(video_path)
|
21 |
+
frames = []
|
22 |
+
ball_positions = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
|
|
24 |
while cap.isOpened():
|
|
|
25 |
ret, frame = cap.read()
|
26 |
if not ret:
|
27 |
break
|
28 |
+
frames.append(frame.copy()) # Store original frame
|
29 |
+
# Detect ball using the trained YOLOv8n model
|
30 |
+
results = model.predict(frame, conf=0.5) # Adjust confidence threshold if needed
|
31 |
+
for detection in results[0].boxes:
|
32 |
+
if detection.cls == 0: # Assuming class 0 is the ball
|
33 |
+
x1, y1, x2, y2 = detection.xyxy[0].cpu().numpy()
|
34 |
+
ball_positions.append([(x1 + x2) / 2, (y1 + y2) / 2])
|
35 |
+
# Draw bounding box on frame for visualization
|
36 |
+
cv2.rectangle(frame, (int(x1), int(y1)), (int(x2), int(y2)), (0, 255, 0), 2)
|
37 |
+
frames[-1] = frame # Update frame with bounding box
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
cap.release()
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
|
40 |
+
return frames, ball_positions
|
|
|
|
|
|
|
41 |
|
42 |
+
def estimate_trajectory(ball_positions, frames):
|
43 |
+
# Simplified physics-based trajectory projection
|
44 |
+
if len(ball_positions) < 2:
|
45 |
+
return None, None
|
46 |
+
# Extract x, y coordinates
|
47 |
+
x_coords = [pos[0] for pos in ball_positions]
|
48 |
+
y_coords = [pos[1] for pos in ball_positions]
|
49 |
+
times = np.arange(len(ball_positions)) / FRAME_RATE
|
50 |
|
51 |
+
# Interpolate to smooth trajectory
|
52 |
try:
|
53 |
+
fx = interp1d(times, x_coords, kind='linear', fill_value="extrapolate")
|
54 |
+
fy = interp1d(times, y_coords, kind='quadratic', fill_value="extrapolate")
|
55 |
except:
|
56 |
+
return None, None
|
57 |
+
|
58 |
+
# Project trajectory forward (0.5 seconds post-impact)
|
59 |
+
t_future = np.linspace(times[-1], times[-1] + 0.5, 10)
|
60 |
+
x_future = fx(t_future)
|
61 |
+
y_future = fy(t_future)
|
62 |
+
|
63 |
+
return list(zip(x_future, y_future)), t_future
|
64 |
+
|
65 |
+
def lbw_decision(ball_positions, trajectory, frames):
|
66 |
+
# Simplified LBW logic
|
67 |
+
if not trajectory or len(ball_positions) < 2:
|
68 |
+
return "Not enough data", None
|
69 |
+
|
70 |
+
# Assume stumps are at the bottom center of the frame (calibration needed)
|
71 |
+
frame_height, frame_width = frames[0].shape[:2]
|
72 |
+
stumps_x = frame_width / 2
|
73 |
+
stumps_y = frame_height * 0.9 # Approximate stumps position
|
74 |
+
stumps_width_pixels = frame_width * (STUMPS_WIDTH / 3.0) # Assume 3m pitch width
|
75 |
+
|
76 |
+
# Check pitching point (first detected position)
|
77 |
+
pitch_x, pitch_y = ball_positions[0]
|
78 |
+
if pitch_x < stumps_x - stumps_width_pixels / 2 or pitch_x > stumps_x + stumps_width_pixels / 2:
|
79 |
+
return "Not Out (Pitched outside line)", None
|
80 |
+
|
81 |
+
# Check impact point (last detected position)
|
82 |
+
impact_x, impact_y = ball_positions[-1]
|
83 |
+
if impact_x < stumps_x - stumps_width_pixels / 2 or impact_x > stumps_x + stumps_width_pixels / 2:
|
84 |
+
return "Not Out (Impact outside line)", None
|
85 |
+
|
86 |
+
# Check trajectory hitting stumps
|
87 |
+
for x, y in trajectory:
|
88 |
+
if abs(x - stumps_x) < stumps_width_pixels / 2 and abs(y - stumps_y) < frame_height * 0.1:
|
89 |
+
return "Out", trajectory
|
90 |
+
return "Not Out (Missing stumps)", trajectory
|
91 |
+
|
92 |
+
def generate_slow_motion(frames, trajectory, output_path):
|
93 |
+
# Generate slow-motion video with ball detection and trajectory overlay
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
95 |
+
out = cv2.VideoWriter(output_path, fourcc, FRAME_RATE / 2, (frames[0].shape[1], frames[0].shape[0]))
|
|
|
|
|
|
|
96 |
|
97 |
+
for frame in frames:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
if trajectory:
|
99 |
+
for x, y in trajectory:
|
100 |
+
cv2.circle(frame, (int(x), int(y)), 5, (255, 0, 0), -1) # Blue dots for trajectory
|
101 |
+
out.write(frame)
|
102 |
+
out.write(frame) # Duplicate frames for slow-motion effect
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
103 |
out.release()
|
104 |
+
return output_path
|
105 |
+
|
106 |
+
def drs_review(video):
|
107 |
+
# Process video and generate DRS output
|
108 |
+
if not os.path.exists(video):
|
109 |
+
return "Error: Video file not found", None
|
110 |
+
frames, ball_positions = process_video(video)
|
111 |
+
trajectory, _ = estimate_trajectory(ball_positions, frames)
|
112 |
+
decision, trajectory = lbw_decision(ball_positions, trajectory, frames)
|
113 |
|
114 |
+
# Generate slow-motion replay
|
115 |
+
output_path = f"output_{uuid.uuid4()}.mp4"
|
116 |
+
slow_motion_path = generate_slow_motion(frames, trajectory, output_path)
|
117 |
|
118 |
+
return decision, slow_motion_path
|
119 |
|
120 |
# Gradio interface
|
121 |
+
iface = gr.Interface(
|
122 |
+
fn=drs_review,
|
123 |
+
inputs=gr.Video(label="Upload Video Clip"),
|
124 |
+
outputs=[
|
125 |
+
gr.Textbox(label="DRS Decision"),
|
126 |
+
gr.Video(label="Slow-Motion Replay with Ball Detection and Trajectory")
|
127 |
+
],
|
128 |
+
title="AI-Powered DRS for LBW in Local Cricket",
|
129 |
+
description="Upload a video clip of a cricket delivery to get an LBW decision and slow-motion replay showing ball detection and trajectory."
|
130 |
+
)
|
131 |
|
132 |
if __name__ == "__main__":
|
133 |
+
iface.launch()
|