def __init__(self, video_name):
self.vid = PreProcess()
self.vid.read_video(video_name)
self.flow = OptFlow()
self.vid.setVideoDimension(100)
self.index = 0
self.binary_mags = []
def getViolentFlow(self):
vid = PreProcess()
vid.read_video(self.video_name)
flow = OptFlow()
vid.setVideoDimension(100)
index = 0
temp_flows = []
for each_frame_index in range(3,vid.total_frames - vid.FRAME_GAP - 5,vid.FRAME_GAP):
PREV_F = vid.getFrameFromIndex(each_frame_index)
CURRENT_F = vid.getFrameFromIndex(each_frame_index + vid.MOVEMENT_INTERVAL)
NEXT_F = vid.getFrameFromIndex(each_frame_index + (2 * vid.MOVEMENT_INTERVAL))
PREV_F = vid.resize_frame(PREV_F)
CURRENT_F = vid.resize_frame(CURRENT_F)
NEXT_F = vid.resize_frame(NEXT_F)
(vx1,vy1,w1) = flow.sorFlow(PREV_F,CURRENT_F)
(vx2,vy2,w2) = flow.sorFlow(CURRENT_F,NEXT_F)
m1 = flow.getFlowMagnitude(vx1,vy1)
index = index + 1
m2 = flow.getFlowMagnitude(vx2,vy2)
# all_zeros1 = not np.any(m1)
# all_zeros2 = not np.any(m2)
# if (all_zeros1 and all_zeros2):
# continue
# index = index + 1
change_mag = abs(m2-m1)
binary_mag = np.ones(change_mag.shape,dtype=np.float64)
threshold = np.mean(change_mag , dtype=np.float64)
temp_flows.append(np.where(change_mag < threshold,0,binary_mag))
flow_video = np.zeros(change_mag.shape,dtype=np.float64)
for each_flow in temp_flows:
flow_video = flow_video + each_flow
flow_video = flow_video / index
#original
# self.height = flow_video.shape[0]
# self.width = flow_video.shape[1]
# self.B_height = int(math.floor(float(self.height - 11)/4))
# self.B_width = int(math.floor(float(self.width - 11)/4))
# testing
self.height = flow_video.shape[0]
self.width = flow_video.shape[1]
self.B_height = int(math.floor(float(self.height - 10)/5))
self.B_width = int(math.floor(float(self.width - 10)/5))
# print self.height , self.width , self.B_height , self.B_width
return flow_video
def __init__(self):
json_file = open('model_100.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
self.model = model_from_json(loaded_model_json)
self.model.load_weights("model_100.h5")
print 'loaded model from disk'
self.vid = PreProcess()
self.vid.setVideoDimension(100)
self.flow = OptFlow()
self.height = 0
self.width = 0
self.B_height = 0
self.B_width = 0
self.index = 0
self.temp_flows = []
self.bins = np.arange(0.0, 1.05, 0.05, dtype=np.float64)
def getViolentFlow(video_name):
vid = PreProcess()
vid.read_video(video_name)
flow = OptFlow()
vid.setVideoDimension(100)
index = 0
temp_flows = []
for each_frame_index in range(3,vid.total_frames - vid.FRAME_GAP - 5,vid.FRAME_GAP):
PREV_F = vid.getFrameFromIndex(each_frame_index)
CURRENT_F = vid.getFrameFromIndex(each_frame_index + vid.MOVEMENT_INTERVAL)
NEXT_F = vid.getFrameFromIndex(each_frame_index + (2 * vid.MOVEMENT_INTERVAL))
PREV_F = vid.resize_frame(PREV_F)
CURRENT_F = vid.resize_frame(CURRENT_F)
NEXT_F = vid.resize_frame(NEXT_F)
(vx1,vy1,w1) = flow.sorFlow(PREV_F,CURRENT_F)
(vx2,vy2,w2) = flow.sorFlow(CURRENT_F,NEXT_F)
m1 = flow.getFlowMagnitude(vx1,vy1)
index = index + 1
m2 = flow.getFlowMagnitude(vx2,vy2)
# print m1
# print m2
change_mag = abs(m2-m1)
binary_mag = np.ones(change_mag.shape,dtype=np.float64)
threshold = np.mean(change_mag , dtype=np.float64)
temp_flows.append(np.where(change_mag < threshold,0,binary_mag))
flow_video = np.zeros(change_mag.shape,dtype=np.float64)
for each_flow in temp_flows:
flow_video = flow_video + each_flow
flow_video = flow_video / index
return flow_video
import cv2
import numpy as np
from VideoProcess import PreProcess
from OpticalFlow import OptFlow
import math
vid = PreProcess()
vid.read_video('vio_1.avi')
flow = OptFlow()
vid.setVideoDimension(100)
index = 0
all_hists1 = []
all_hists2 = []
def binData(X, bins, mag):
hist = [0.0] * 9
inds = np.digitize(X, bins)
for i in range(X.size):
hist[inds[i][0] - 1] = hist[inds[i][0] - 1] + mag[i][0]
return hist
for each_frame_index in range(3, vid.total_frames - vid.FRAME_GAP - 5,
vid.FRAME_GAP):
PREV_F = vid.getFrameFromIndex(each_frame_index)
CURRENT_F = vid.getFrameFromIndex(each_frame_index + vid.MOVEMENT_INTERVAL)
NEXT_F = vid.getFrameFromIndex(each_frame_index +
(2 * vid.MOVEMENT_INTERVAL))
class ContinousSurv2:
def __init__(self):
json_file = open('model_100.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
self.model = model_from_json(loaded_model_json)
self.model.load_weights("model_100.h5")
print 'loaded model from disk'
self.vid = PreProcess()
self.vid.setVideoDimension(100)
self.flow = OptFlow()
self.height = 0
self.width = 0
self.B_height = 0
self.B_width = 0
self.index = 0
self.temp_flows = []
self.bins = np.arange(0.0, 1.05, 0.05, dtype=np.float64)
def setVideoName(self, video_name):
self.vid.read_video(video_name)
def histc(self, X, bins):
map_to_bins = np.digitize(X, bins)
r = np.zeros(bins.shape, dtype=np.float64)
for i in map_to_bins:
r[i - 1] += 1
return r
def getBlockHist(self, flow_video):
flow_vec = np.reshape(flow_video,
(flow_video.shape[0] * flow_video.shape[1], 1))
count_of_bins = self.histc(flow_vec, self.bins)
return count_of_bins / np.sum(count_of_bins)
def getFrameHist(self, flow_video_size):
flow_video = np.zeros(flow_video_size, dtype=np.float64)
for each_flow in self.temp_flows:
flow_video = flow_video + each_flow
flow_video = flow_video / self.index
self.index = 0
self.temp_flows = []
self.height = flow_video.shape[0]
self.width = flow_video.shape[1]
self.B_height = int(math.floor((self.height - 11) / 4))
self.B_width = int(math.floor((self.width - 11) / 4))
frame_hist = []
for y in range(6, self.height - self.B_height - 4, self.B_height):
for x in range(6, self.width - self.B_width - 4, self.B_width):
block_hist = self.getBlockHist(flow_video[y:y + self.B_height,
x:x + self.B_width])
frame_hist = np.append(frame_hist, block_hist, axis=0)
return frame_hist
def doSurveillanceFromVideo(self):
video_time_start = time.time()
FPS = round(self.vid.getFPS())
print 'FPS is : ' + str(FPS)
while True:
frames = self.vid.getFramesFromVideoSource()
PREV_F = frames[0]
CURRENT_F = frames[1]
NEXT_F = frames[2]
frame_number = frames[3]
PREV_F = self.vid.resize_frame(PREV_F)
CURRENT_F = self.vid.resize_frame(CURRENT_F)
NEXT_F = self.vid.resize_frame(NEXT_F)
(vx1, vy1, w1) = self.flow.sorFlow(PREV_F, CURRENT_F)
(vx2, vy2, w2) = self.flow.sorFlow(CURRENT_F, NEXT_F)
m1 = self.flow.getFlowMagnitude(vx1, vy1)
self.index = self.index + 1
m2 = self.flow.getFlowMagnitude(vx2, vy2)
change_mag = abs(m2 - m1)
binary_mag = np.ones(change_mag.shape, dtype=np.float64)
threshold = np.mean(change_mag, dtype=np.float64)
self.temp_flows.append(
np.where(change_mag < threshold, 0, binary_mag))
if self.index >= int(FPS / 3):
vif = self.getFrameHist(CURRENT_F.shape)
X_frame = np.empty((0, 336))
vif = np.reshape(vif, (-1, vif.shape[0]))
X_frame = np.vstack((X_frame, vif))
pred = self.model.predict(X_frame)
pred = round(pred[0][0])
if pred == 1:
time_violence = float(frame_number) / self.vid.fps
print 'violent --- ' + str(
int(time_violence)
) + ' seconds , processing time : ' + str(time.time() -
video_time_start)
def doSurveillanceFromCamera(self):
start_time = time.time()
self.vid.useCamera()
FPS = round(self.vid.getFPS())
print 'FPS is :' + str(FPS)
while True:
frames = self.vid.getFramesFromCameraSource()
PREV_F = frames[0]
CURRENT_F = frames[1]
NEXT_F = frames[2]
time_now = frames[3]
PREV_F = self.vid.resize_frame(PREV_F)
CURRENT_F = self.vid.resize_frame(CURRENT_F)
NEXT_F = self.vid.resize_frame(NEXT_F)
(vx1, vy1, w1) = self.flow.sorFlow(PREV_F, CURRENT_F)
(vx2, vy2, w2) = self.flow.sorFlow(CURRENT_F, NEXT_F)
m1 = self.flow.getFlowMagnitude(vx1, vy1)
self.index = self.index + 1
m2 = self.flow.getFlowMagnitude(vx2, vy2)
change_mag = abs(m2 - m1)
binary_mag = np.ones(change_mag.shape, dtype=np.float64)
threshold = np.mean(change_mag, dtype=np.float64)
self.temp_flows.append(
np.where(change_mag < threshold, 0, binary_mag))
if self.index >= int(FPS / 3):
vif = self.getFrameHist(CURRENT_F.shape)
X_frame = np.empty((0, 336))
vif = np.reshape(vif, (-1, vif.shape[0]))
X_frame = np.vstack((X_frame, vif))
pred = self.model.predict(X_frame)
print pred, time_now - start_time
pred = round(pred[0][0])
if pred == 1:
print 'violent --- ' + str(time_now - start_time)
import cv2
from VideoProcess import PreProcess
from OpticalFlow import OptFlow
vid = PreProcess()
flow = OptFlow()
vid.useCamera()
vid.setVideoDimension(300)
while True:
frame1 = vid.getFramesFromSource()
frame2 = vid.getFramesFromSource()
(vx, vy, w) = flow.sorFlow(frame1, frame2)
cv2.imshow('camera', frame2)
cv2.imshow('vy', vy)
cv2.imshow('vx', vx)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
class OvifCalc:
def __init__(self, video_name):
self.vid = PreProcess()
self.vid.read_video(video_name)
self.flow = OptFlow()
self.vid.setVideoDimension(100)
self.index = 0
self.binary_mags = []
def binData(self, X, bins, mag):
hist = [0.0] * 9
inds = np.digitize(X, bins)
for i in range(X.size):
hist[inds[i][0] - 1] = hist[inds[i][0] - 1] + mag[i][0]
return hist
def getBinaryMag(self, hist1, hist2):
hist1 = np.array(hist1)
hist2 = np.array(hist2)
change_mag = abs(hist1 - hist2)
threshold = np.mean(change_mag, dtype=np.float64)
binary_mag = np.ones(change_mag.shape, dtype=np.float64)
binary_mag = np.where(change_mag < threshold, 0, binary_mag)
return binary_mag
def getOvifFeature(self):
for each_frame_index in range(
3, self.vid.total_frames - self.vid.FRAME_GAP - 5,
self.vid.FRAME_GAP):
PREV_F = self.vid.getFrameFromIndex(each_frame_index)
CURRENT_F = self.vid.getFrameFromIndex(each_frame_index +
self.vid.MOVEMENT_INTERVAL)
NEXT_F = self.vid.getFrameFromIndex(each_frame_index +
(2 *
self.vid.MOVEMENT_INTERVAL))
PREV_F = self.vid.resize_frame(PREV_F)
CURRENT_F = self.vid.resize_frame(CURRENT_F)
NEXT_F = self.vid.resize_frame(NEXT_F)
(vx1, vy1, w1) = self.flow.sorFlow(PREV_F, CURRENT_F)
(vx2, vy2, w2) = self.flow.sorFlow(CURRENT_F, NEXT_F)
m1 = self.flow.getFlowMagnitude(vx1, vy1)
self.index = self.index + 1
m2 = self.flow.getFlowMagnitude(vx2, vy2)
all_zeros1 = not np.any(vx1)
all_zeros2 = not np.any(vx2)
all_hists1 = []
all_hists2 = []
if not (all_zeros1 or all_zeros2):
phi1 = np.arctan(vy1 / vx1)
phi2 = np.arctan(vy2 / vx2)
data_block_size = (m1.shape[0] * m1.shape[1]) / (4 * 4)
m1 = np.reshape(m1, (m1.shape[0] * m1.shape[1], 1))
m2 = np.reshape(m2, (m2.shape[0] * m2.shape[1], 1))
phi1 = np.reshape(phi1, (phi1.shape[0] * phi1.shape[1], 1))
phi2 = np.reshape(phi2, (phi2.shape[0] * phi2.shape[1], 1))
hist_bins = np.arange(9.) * (2 * np.pi) / 9
hist_bins = np.degrees(hist_bins)
phi1 = np.rad2deg(phi1)
phi2 = np.rad2deg(phi2)
phi1 = np.where(phi1 < 0, 360 + phi1, phi1)
phi2 = np.where(phi2 < 0, 360 + phi2, phi2)
for k in range(0, (4 * 4)):
data_to_be_binned1 = []
data_to_be_binned2 = []
if k != (4 * 4) - 1:
data_to_be_binned1 = m1[data_block_size *
k:data_block_size * (k + 1)]
temp_phi1 = phi1[data_block_size * k:data_block_size *
(k + 1)]
data_to_be_binned2 = m2[data_block_size *
k:data_block_size * (k + 1)]
temp_phi2 = phi2[data_block_size * k:data_block_size *
(k + 1)]
else:
data_to_be_binned1 = m1[data_block_size * k:]
temp_phi1 = phi1[data_block_size * k:]
data_to_be_binned2 = m2[data_block_size * k:]
temp_phi2 = phi2[data_block_size * k:]
hist1 = self.binData(temp_phi1, hist_bins,
data_to_be_binned1)
hist2 = self.binData(temp_phi2, hist_bins,
data_to_be_binned2)
all_hists1 = all_hists1 + hist1
all_hists2 = all_hists2 + hist2
binary_mag = self.getBinaryMag(all_hists1, all_hists2)
self.binary_mags.append(binary_mag)
else:
binary_mag = np.zeros(((4 * 4) * (9), 1), dtype=np.float64)
self.binary_mags.append(binary_mag)
return self.getAverageBinary()
def getAverageBinary(self):
ovif_video = np.zeros(self.binary_mags[0].shape, dtype=np.float64)
for each_ovif in self.binary_mags:
ovif_video = ovif_video + each_ovif
return ovif_video / self.index
def writeFeatureToFile(self, filename):
np.savetxt(filename, self.getOvifFeature(), delimiter=',')