|
|
import cv2
|
|
|
import torch
|
|
|
import torch.nn as nn
|
|
|
import torch.optim as optim
|
|
|
import numpy as np
|
|
|
|
|
|
|
|
|
device = torch.device("cpu")
|
|
|
|
|
|
label_map = {"Idle": 0, "Normal": 1, "Erratic": 2}
|
|
|
reverse_label = {v: k for k, v in label_map.items()}
|
|
|
|
|
|
class BehaviorAI(nn.Module):
|
|
|
def __init__(self):
|
|
|
super().__init__()
|
|
|
self.model = nn.Sequential(
|
|
|
nn.Linear(4, 16),
|
|
|
nn.ReLU(),
|
|
|
nn.Linear(16, 8),
|
|
|
nn.ReLU(),
|
|
|
nn.Linear(8, 3)
|
|
|
)
|
|
|
self.loss_fn = nn.CrossEntropyLoss()
|
|
|
self.optimizer = optim.Adam(self.model.parameters(), lr=0.001)
|
|
|
|
|
|
def forward(self, x):
|
|
|
return self.model(x)
|
|
|
|
|
|
def predict_behavior(self, features):
|
|
|
self.model.eval()
|
|
|
with torch.no_grad():
|
|
|
x = torch.tensor([features], dtype=torch.float32).to(device)
|
|
|
logits = self.model(x)
|
|
|
pred = torch.argmax(logits, dim=-1).item()
|
|
|
return reverse_label[pred]
|
|
|
|
|
|
def learn_from(self, features, label):
|
|
|
self.model.train()
|
|
|
x = torch.tensor([features], dtype=torch.float32).to(device)
|
|
|
y = torch.tensor([label_map[label]], dtype=torch.long).to(device)
|
|
|
logits = self.model(x)
|
|
|
loss = self.loss_fn(logits, y)
|
|
|
self.optimizer.zero_grad()
|
|
|
loss.backward()
|
|
|
self.optimizer.step()
|
|
|
|
|
|
|
|
|
def extract_features(trace):
|
|
|
if len(trace) < 2:
|
|
|
return [0, 0, 0, 0]
|
|
|
|
|
|
dx = trace[-1][0] - trace[0][0]
|
|
|
dy = trace[-1][1] - trace[0][1]
|
|
|
speeds = []
|
|
|
directions = []
|
|
|
|
|
|
for i in range(1, len(trace)):
|
|
|
x1, y1 = trace[i-1]
|
|
|
x2, y2 = trace[i]
|
|
|
dist = np.linalg.norm([x2 - x1, y2 - y1])
|
|
|
speeds.append(dist)
|
|
|
directions.append(np.arctan2(y2 - y1, x2 - x1))
|
|
|
|
|
|
avg_speed = np.mean(speeds)
|
|
|
direction_changes = np.sum(np.abs(np.diff(directions)))
|
|
|
return [dx, dy, avg_speed, direction_changes]
|
|
|
|
|
|
|
|
|
cap = cv2.VideoCapture(0)
|
|
|
|
|
|
bg_subtractor = cv2.createBackgroundSubtractorMOG2()
|
|
|
traces = {}
|
|
|
next_id = 0
|
|
|
ai = BehaviorAI()
|
|
|
|
|
|
while True:
|
|
|
ret, frame = cap.read()
|
|
|
if not ret:
|
|
|
break
|
|
|
|
|
|
fgmask = bg_subtractor.apply(frame)
|
|
|
contours, _ = cv2.findContours(fgmask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
|
|
|
|
|
current_positions = []
|
|
|
|
|
|
for cnt in contours:
|
|
|
if cv2.contourArea(cnt) < 500:
|
|
|
continue
|
|
|
|
|
|
x, y, w, h = cv2.boundingRect(cnt)
|
|
|
cx, cy = x + w // 2, y + h // 2
|
|
|
current_positions.append((cx, cy))
|
|
|
cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
|
|
|
|
|
|
new_traces = {}
|
|
|
matched_ids = set()
|
|
|
|
|
|
for cx, cy in current_positions:
|
|
|
min_dist = float('inf')
|
|
|
matched_id = None
|
|
|
for id, trace in traces.items():
|
|
|
if len(trace) == 0:
|
|
|
continue
|
|
|
prev_x, prev_y = trace[-1]
|
|
|
dist = np.linalg.norm([cx - prev_x, cy - prev_y])
|
|
|
if dist < 50 and id not in matched_ids:
|
|
|
min_dist = dist
|
|
|
matched_id = id
|
|
|
|
|
|
if matched_id is None:
|
|
|
matched_id = next_id
|
|
|
next_id += 1
|
|
|
new_traces[matched_id] = []
|
|
|
|
|
|
else:
|
|
|
new_traces[matched_id] = traces[matched_id]
|
|
|
|
|
|
new_traces[matched_id].append((cx, cy))
|
|
|
matched_ids.add(matched_id)
|
|
|
|
|
|
traces = new_traces
|
|
|
|
|
|
for id, trace in traces.items():
|
|
|
if len(trace) >= 2:
|
|
|
for i in range(1, len(trace)):
|
|
|
cv2.line(frame, trace[i-1], trace[i], (255, 0, 0), 2)
|
|
|
|
|
|
features = extract_features(trace)
|
|
|
behavior = ai.predict_behavior(features)
|
|
|
|
|
|
if len(trace) >= 10:
|
|
|
if features[2] < 2:
|
|
|
label = "Idle"
|
|
|
elif features[3] > 4:
|
|
|
label = "Erratic"
|
|
|
else:
|
|
|
label = "Normal"
|
|
|
ai.learn_from(features, label)
|
|
|
|
|
|
cv2.putText(frame, f"ID:{id} AI:{behavior}", trace[-1], cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 255), 1)
|
|
|
|
|
|
cv2.imshow("Real-Time Tracker with AI", frame)
|
|
|
if cv2.waitKey(1) == 27:
|
|
|
break
|
|
|
|
|
|
cap.release()
|
|
|
cv2.destroyAllWindows()
|
|
|
|
