Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import cv2
|
3 |
+
import tempfile
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
from collections import deque
|
7 |
+
from transformers import AutoFeatureExtractor, AutoModelForVideoClassification
|
8 |
+
from streamlit_webrtc import webrtc_streamer, VideoTransformerBase, RTCConfiguration
|
9 |
+
|
10 |
+
# Constants
|
11 |
+
NUM_FRAMES = 16
|
12 |
+
MODEL_NAME = "jatinmehra/Accident-Detection-using-Dashcam"
|
13 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
14 |
+
|
15 |
+
@st.cache_resource
|
16 |
+
def load_model_and_extractor():
|
17 |
+
extractor = AutoFeatureExtractor.from_pretrained("facebook/timesformer-base-finetuned-k400")
|
18 |
+
model = AutoModelForVideoClassification.from_pretrained(
|
19 |
+
MODEL_NAME,
|
20 |
+
num_labels=2,
|
21 |
+
ignore_mismatched_sizes=True
|
22 |
+
).to(DEVICE)
|
23 |
+
model.eval()
|
24 |
+
return extractor, model
|
25 |
+
|
26 |
+
extractor, model = load_model_and_extractor()
|
27 |
+
|
28 |
+
st.title("Dashcam Accident Predictor")
|
29 |
+
st.write("**higher score = higher accident probability**")
|
30 |
+
|
31 |
+
# Function to run inference on a saved video file
|
32 |
+
|
33 |
+
def run_inference_on_video(video_path):
|
34 |
+
cap = cv2.VideoCapture(video_path)
|
35 |
+
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
36 |
+
fps = cap.get(cv2.CAP_PROP_FPS) or 30
|
37 |
+
if total_frames <= 0:
|
38 |
+
st.error("Failed to read video frames.")
|
39 |
+
return None
|
40 |
+
|
41 |
+
# Uniform sampling
|
42 |
+
indices = np.linspace(0, total_frames-1, NUM_FRAMES, dtype=int)
|
43 |
+
frames = []
|
44 |
+
for idx in indices:
|
45 |
+
cap.set(cv2.CAP_PROP_POS_FRAMES, int(idx))
|
46 |
+
ret, frame = cap.read()
|
47 |
+
if not ret:
|
48 |
+
frames.append(np.zeros((224,224,3), dtype=np.uint8))
|
49 |
+
else:
|
50 |
+
rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
51 |
+
resized = cv2.resize(rgb, (224,224))
|
52 |
+
frames.append(resized)
|
53 |
+
cap.release()
|
54 |
+
|
55 |
+
# Preprocess and predict
|
56 |
+
inputs = extractor(frames, return_tensors="pt")
|
57 |
+
pixel_values = inputs['pixel_values'].to(DEVICE)
|
58 |
+
with torch.no_grad():
|
59 |
+
outputs = model(pixel_values=pixel_values).logits
|
60 |
+
prob = torch.softmax(outputs, dim=1)[0,1].item()
|
61 |
+
return prob
|
62 |
+
|
63 |
+
# UI Selection
|
64 |
+
source = st.radio("Choose input source", ("Upload Video", "Webcam"))
|
65 |
+
|
66 |
+
if source == "Upload Video":
|
67 |
+
uploaded_file = st.file_uploader("Upload a video file", type=["mp4", "avi", "mov"])
|
68 |
+
if uploaded_file is not None:
|
69 |
+
tfile = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
|
70 |
+
tfile.write(uploaded_file.read())
|
71 |
+
st.video(uploaded_file)
|
72 |
+
st.write("Running inference...")
|
73 |
+
score = run_inference_on_video(tfile.name)
|
74 |
+
if score is not None:
|
75 |
+
st.success(f"Accident probability: {score:.2f}")
|
76 |
+
|
77 |
+
else:
|
78 |
+
# Webcam stream processing
|
79 |
+
class AcciTransformer(VideoTransformerBase):
|
80 |
+
def __init__(self):
|
81 |
+
self.buffer = deque(maxlen=NUM_FRAMES)
|
82 |
+
|
83 |
+
def transform(self, frame):
|
84 |
+
img = frame.to_ndarray(format="bgr24")
|
85 |
+
rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
86 |
+
resized = cv2.resize(rgb, (224,224))
|
87 |
+
self.buffer.append(resized)
|
88 |
+
if len(self.buffer) == NUM_FRAMES:
|
89 |
+
inputs = extractor(list(self.buffer), return_tensors="pt")
|
90 |
+
pixel_values = inputs['pixel_values'].to(DEVICE)
|
91 |
+
with torch.no_grad():
|
92 |
+
outputs = model(pixel_values=pixel_values).logits
|
93 |
+
prob = torch.softmax(outputs, dim=1)[0,1].item()
|
94 |
+
cv2.putText(img, f"Prob: {prob:.2f}", (10,30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0,0,255), 2)
|
95 |
+
return img
|
96 |
+
|
97 |
+
webrtc_streamer(
|
98 |
+
key="dashcam-webcam",
|
99 |
+
mode="recv",
|
100 |
+
rtc_configuration=RTCConfiguration({
|
101 |
+
"iceServers": [{"urls": ["stun:stun.l.google.com:19302"]}]
|
102 |
+
}),
|
103 |
+
video_transformer_factory=AcciTransformer
|
104 |
+
)
|