Update app.py
Browse files
app.py
CHANGED
@@ -14,151 +14,107 @@ warnings.filterwarnings('ignore', category=FutureWarning)
|
|
14 |
|
15 |
class RobustSafetyMonitor:
|
16 |
def __init__(self):
|
17 |
-
"""Initialize the
|
18 |
self.client = Groq()
|
19 |
-
self.model_name = "llama-3.2-11b-vision-preview"
|
20 |
self.max_image_size = (800, 800)
|
21 |
self.colors = [(0, 0, 255), (255, 0, 0), (0, 255, 0), (255, 255, 0), (255, 0, 255)]
|
22 |
|
23 |
-
# Load YOLOv5 model
|
24 |
self.yolo_model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True)
|
25 |
-
|
26 |
-
|
|
|
|
|
27 |
self.yolo_model.cpu()
|
28 |
self.yolo_model.eval()
|
29 |
|
30 |
-
def preprocess_image(self, frame: np.ndarray) -> np.ndarray:
|
31 |
-
"""Process image for analysis."""
|
32 |
-
if frame is None:
|
33 |
-
raise ValueError("No image provided")
|
34 |
-
|
35 |
-
if len(frame.shape) == 2:
|
36 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_GRAY2RGB)
|
37 |
-
elif len(frame.shape) == 3 and frame.shape[2] == 4:
|
38 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_RGBA2RGB)
|
39 |
-
|
40 |
-
return self.resize_image(frame)
|
41 |
-
|
42 |
-
def resize_image(self, image: np.ndarray) -> np.ndarray:
|
43 |
-
"""Resize image while maintaining aspect ratio."""
|
44 |
-
height, width = image.shape[:2]
|
45 |
-
if height > self.max_image_size[1] or width > self.max_image_size[0]:
|
46 |
-
aspect = width / height
|
47 |
-
if width > height:
|
48 |
-
new_width = self.max_image_size[0]
|
49 |
-
new_height = int(new_width / aspect)
|
50 |
-
else:
|
51 |
-
new_height = self.max_image_size[1]
|
52 |
-
new_width = int(new_height * aspect)
|
53 |
-
return cv2.resize(image, (new_width, new_height), interpolation=cv2.INTER_AREA)
|
54 |
-
return image
|
55 |
-
|
56 |
-
def encode_image(self, frame: np.ndarray) -> str:
|
57 |
-
"""Convert image to base64 encoding with proper formatting."""
|
58 |
-
try:
|
59 |
-
frame_pil = PILImage.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
|
60 |
-
buffered = io.BytesIO()
|
61 |
-
frame_pil.save(buffered, format="JPEG", quality=95)
|
62 |
-
img_base64 = base64.b64encode(buffered.getvalue()).decode('utf-8')
|
63 |
-
return f"data:image/jpeg;base64,{img_base64}"
|
64 |
-
except Exception as e:
|
65 |
-
raise ValueError(f"Error encoding image: {str(e)}")
|
66 |
-
|
67 |
def detect_objects(self, frame: np.ndarray) -> Tuple[np.ndarray, Dict]:
|
68 |
-
"""
|
69 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
with torch.no_grad():
|
71 |
-
results = self.yolo_model(frame)
|
|
|
|
|
72 |
bbox_data = results.xyxy[0].cpu().numpy()
|
73 |
labels = results.names
|
74 |
-
return bbox_data, labels
|
75 |
-
except Exception as e:
|
76 |
-
raise ValueError(f"Error detecting objects: {str(e)}")
|
77 |
-
|
78 |
-
def analyze_frame(self, frame: np.ndarray) -> Tuple[List[Dict], str]:
|
79 |
-
"""Perform safety analysis on the frame using Llama Vision."""
|
80 |
-
if frame is None:
|
81 |
-
return [], "No frame received"
|
82 |
-
|
83 |
-
try:
|
84 |
-
frame = self.preprocess_image(frame)
|
85 |
-
image_base64 = self.encode_image(frame)
|
86 |
-
|
87 |
-
completion = self.client.chat.completions.create(
|
88 |
-
model=self.model_name,
|
89 |
-
messages=[
|
90 |
-
{
|
91 |
-
"role": "user",
|
92 |
-
"content": [
|
93 |
-
{
|
94 |
-
"type": "text",
|
95 |
-
"text": """Analyze this workplace image and identify any potential safety risks.
|
96 |
-
List each risk on a new line starting with 'Risk:'.
|
97 |
-
Format: Risk: [Object/Area] - [Description of hazard]"""
|
98 |
-
},
|
99 |
-
{
|
100 |
-
"type": "image_url",
|
101 |
-
"image_url": {
|
102 |
-
"url": image_base64
|
103 |
-
}
|
104 |
-
}
|
105 |
-
]
|
106 |
-
}
|
107 |
-
],
|
108 |
-
temperature=0.7,
|
109 |
-
max_tokens=1024,
|
110 |
-
stream=False
|
111 |
-
)
|
112 |
|
113 |
-
#
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
except Exception as e:
|
123 |
-
print(f"
|
124 |
-
return [],
|
125 |
|
126 |
def draw_bounding_boxes(self, image: np.ndarray, bboxes: np.ndarray,
|
127 |
labels: Dict, safety_issues: List[Dict]) -> np.ndarray:
|
128 |
-
"""
|
129 |
image_copy = image.copy()
|
130 |
font = cv2.FONT_HERSHEY_SIMPLEX
|
131 |
font_scale = 0.5
|
132 |
thickness = 2
|
133 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
134 |
for idx, bbox in enumerate(bboxes):
|
135 |
try:
|
136 |
x1, y1, x2, y2, conf, class_id = bbox
|
137 |
label = labels[int(class_id)]
|
138 |
-
color = self.colors[idx % len(self.colors)]
|
139 |
-
|
140 |
-
# Convert coordinates to integers
|
141 |
-
x1, y1, x2, y2 = map(int, [x1, y1, x2, y2])
|
142 |
|
143 |
-
#
|
144 |
-
|
145 |
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
y_pos = max(y1 - 10, 20)
|
152 |
cv2.putText(image_copy, label_text, (x1, y_pos), font,
|
153 |
-
font_scale,
|
154 |
-
risk_found = True
|
155 |
-
break
|
156 |
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
|
|
162 |
except Exception as e:
|
163 |
print(f"Error drawing box: {str(e)}")
|
164 |
continue
|
|
|
14 |
|
15 |
class RobustSafetyMonitor:
|
16 |
def __init__(self):
|
17 |
+
"""Initialize the safety detection tool with improved configuration."""
|
18 |
self.client = Groq()
|
19 |
+
self.model_name = "llama-3.2-11b-vision-preview"
|
20 |
self.max_image_size = (800, 800)
|
21 |
self.colors = [(0, 0, 255), (255, 0, 0), (0, 255, 0), (255, 255, 0), (255, 0, 255)]
|
22 |
|
23 |
+
# Load YOLOv5 model with improved configuration
|
24 |
self.yolo_model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True)
|
25 |
+
self.yolo_model.conf = 0.25 # Lower confidence threshold for more detections
|
26 |
+
self.yolo_model.iou = 0.45 # Adjusted IOU threshold
|
27 |
+
self.yolo_model.classes = None # Detect all classes
|
28 |
+
self.yolo_model.max_det = 50 # Increased maximum detections
|
29 |
self.yolo_model.cpu()
|
30 |
self.yolo_model.eval()
|
31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
def detect_objects(self, frame: np.ndarray) -> Tuple[np.ndarray, Dict]:
|
33 |
+
"""Enhanced object detection using YOLOv5."""
|
34 |
try:
|
35 |
+
# Ensure proper image format
|
36 |
+
if len(frame.shape) == 2:
|
37 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_GRAY2RGB)
|
38 |
+
elif frame.shape[2] == 4:
|
39 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_RGBA2RGB)
|
40 |
+
|
41 |
+
# Run inference with augmentation
|
42 |
with torch.no_grad():
|
43 |
+
results = self.yolo_model(frame, augment=True) # Enable test-time augmentation
|
44 |
+
|
45 |
+
# Get detections
|
46 |
bbox_data = results.xyxy[0].cpu().numpy()
|
47 |
labels = results.names
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
|
49 |
+
# Filter and process detections
|
50 |
+
processed_boxes = []
|
51 |
+
for box in bbox_data:
|
52 |
+
x1, y1, x2, y2, conf, cls = box
|
53 |
+
# Additional filtering for construction site objects
|
54 |
+
if conf > 0.25: # Keep lower confidence threshold for more detections
|
55 |
+
processed_boxes.append(box)
|
56 |
+
|
57 |
+
return np.array(processed_boxes), labels
|
58 |
except Exception as e:
|
59 |
+
print(f"Error in object detection: {str(e)}")
|
60 |
+
return np.array([]), {}
|
61 |
|
62 |
def draw_bounding_boxes(self, image: np.ndarray, bboxes: np.ndarray,
|
63 |
labels: Dict, safety_issues: List[Dict]) -> np.ndarray:
|
64 |
+
"""Improved bounding box visualization."""
|
65 |
image_copy = image.copy()
|
66 |
font = cv2.FONT_HERSHEY_SIMPLEX
|
67 |
font_scale = 0.5
|
68 |
thickness = 2
|
69 |
|
70 |
+
# Define construction-related keywords for better object association
|
71 |
+
construction_keywords = [
|
72 |
+
'person', 'worker', 'helmet', 'tool', 'machine', 'equipment',
|
73 |
+
'brick', 'block', 'pile', 'stack', 'surface', 'floor', 'ground',
|
74 |
+
'construction', 'building', 'structure'
|
75 |
+
]
|
76 |
+
|
77 |
for idx, bbox in enumerate(bboxes):
|
78 |
try:
|
79 |
x1, y1, x2, y2, conf, class_id = bbox
|
80 |
label = labels[int(class_id)]
|
|
|
|
|
|
|
|
|
81 |
|
82 |
+
# Check if object is construction-related
|
83 |
+
is_relevant = any(keyword in label.lower() for keyword in construction_keywords)
|
84 |
|
85 |
+
if is_relevant or conf > 0.35: # Higher threshold for non-construction objects
|
86 |
+
color = self.colors[idx % len(self.colors)]
|
87 |
+
|
88 |
+
# Convert coordinates to integers
|
89 |
+
x1, y1, x2, y2 = map(int, [x1, y1, x2, y2])
|
90 |
+
|
91 |
+
# Draw thicker bounding box for better visibility
|
92 |
+
cv2.rectangle(image_copy, (x1, y1), (x2, y2), color, thickness)
|
93 |
+
|
94 |
+
# Check for associated safety issues
|
95 |
+
risk_found = False
|
96 |
+
for safety_issue in safety_issues:
|
97 |
+
issue_keywords = safety_issue.get('object', '').lower().split()
|
98 |
+
if any(keyword in label.lower() for keyword in issue_keywords):
|
99 |
+
label_text = f"Risk: {safety_issue.get('description', '')}"
|
100 |
+
y_pos = max(y1 - 10, 20)
|
101 |
+
cv2.putText(image_copy, label_text, (x1, y_pos), font,
|
102 |
+
font_scale, (0, 0, 255), thickness)
|
103 |
+
risk_found = True
|
104 |
+
break
|
105 |
+
|
106 |
+
if not risk_found:
|
107 |
+
label_text = f"{label} {conf:.2f}"
|
108 |
y_pos = max(y1 - 10, 20)
|
109 |
cv2.putText(image_copy, label_text, (x1, y_pos), font,
|
110 |
+
font_scale, color, thickness)
|
|
|
|
|
111 |
|
112 |
+
# Draw additional markers for high-risk areas
|
113 |
+
if conf > 0.5 and any(risk_word in label.lower() for risk_word in
|
114 |
+
['worker', 'person', 'equipment', 'machine']):
|
115 |
+
cv2.circle(image_copy, (int((x1 + x2)/2), int((y1 + y2)/2)),
|
116 |
+
5, (0, 0, 255), -1)
|
117 |
+
|
118 |
except Exception as e:
|
119 |
print(f"Error drawing box: {str(e)}")
|
120 |
continue
|