def run_experiment(cap, disp, timeframes, timesteps, agents):
# We're going to compare ourselves to the agents with the following IDs.
print 'Running experiment...'
util.reset_timer()
agents_to_test = range(319, 331)
# agents_to_test = [48, 49]
true_paths = []
planned_paths = []
one_step_paths = []
# IGP_scores = np.zeros((len(agents_to_test), 2))
# ped_scores = np.zeros((len(agents_to_test), 2))
display.plot_prediction_metrics([], [], [])
for i, agent in enumerate(agents_to_test):
print 'Agent ', agent
ped_path = agents[agent]
path_length = ped_path.shape[0]
start = 3 # Initializes path with start+1 points.
final_path = np.zeros((path_length, 3))
final_path[0:start+1, :] = ped_path[0:start+1,1:4]
one_step_final_path = final_path.copy()
# Run IGP through the path sequence.
for t in range(start, path_length-1):
frame_num = timeframes[int(ped_path[t,0])]
print 'doing frame', frame_num
disp.set_frame(frame_num)
final_path[t+1], one_step_final_path[t+1] = disp.do_frame(agent, final_path[:t+1], False, True)
if cv2.waitKey(1) != -1:
print 'Canceled!'
return
# Compute the final score for both IGP and pedestrian ground truth.
print 'Agent', agent, 'done.'
# start_time = int(ped_path[0,0])
# other_peds = [agents[a] for a in timesteps[start_time] if a != agent]
# other_paths = [util.get_path_at_time(start_time, fullpath)[1][:,1:4] for fullpath in other_peds]
# IGP_scores[i] = util.length_and_safety(final_path, other_paths)
# ped_scores[i] = util.length_and_safety(ped_path[:,1:4], other_paths)
true_paths.append(ped_path[:,1:4])
planned_paths.append(final_path)
one_step_paths.append(one_step_final_path)
pred_errs = util.calc_pred_scores(true_paths, planned_paths, util.prediction_errors)
path_errs = util.calc_pred_scores(true_paths, planned_paths, util.path_errors)
display.plot_prediction_metrics(pred_errs, path_errs, agents_to_test[0:i+1])
one_step_errors = util.calc_pred_scores(true_paths, one_step_paths, util.prediction_errors)
pl.figure(2)
pl.clf()
display.plot_prediction_error('One-Step Prediction Errors', one_step_errors, agents_to_test[0:i+1])
pl.draw()
# results = np.column_stack((agents_to_test, ped_scores, IGP_scores))
pred_results = np.column_stack((agents_to_test, pred_errs.T))
path_results = np.column_stack((agents_to_test, path_errs.T))
# np.savetxt('experiment.txt', results)
np.savetxt('prediction_errors.txt', pred_results)
np.savetxt('path_errors.txt', path_results)
print 'EXPERIMENT COMPLETE.'
util.report_time()
def main():
# Parse command-line arguments.
args = parse_args()
Hfile = os.path.join(args.datadir, "H.txt")
mapfile = os.path.join(args.datadir, "map.png")
obsfile = os.path.join(args.datadir, "obsmat.txt")
destfile = os.path.join(args.datadir, "destinations.txt")
# Parse homography matrix.
H = np.loadtxt(Hfile)
Hinv = np.linalg.inv(H)
# Parse obstacle map.
obs_map = ewap.create_obstacle_map(mapfile)
# Parse pedestrian annotations.
frames, timeframes, timesteps, agents = ewap.parse_annotations(
Hinv, obsfile)
# Parse destinations.
destinations = np.loadtxt(destfile)
# Play the video
seqname = os.path.basename(args.datadir)
cap = cv2.VideoCapture(os.path.join(args.datadir, seqname + ".avi"))
disp = display.Display(cap, Hinv, obs_map, frames, timesteps, agents,
destinations)
disp.model = models.model21
# We can set the exact frame we want to look at:
# disp.set_frame(780) # ETH sequence, beginning
# disp.set_frame(2238) # ETH sequence, agent #48 (1 v 1)
disp.set_frame(11301) # ETH sequence, big crowds both ways
# disp.set_frame(8289) # ETH sequence, agent #175 (1 v crowd)
# disp.set_frame(9867) # ETH sequence, agent #236 (1 v crowd)
# disp.set_frame(8859) # ETH sequence, agent #194 (1 v 1, choosing non-optimal avoidance)
# disp.set_frame(9261) # Hotel sequence, agent #175
# We can also go by minutes & seconds:
# seekpos = 7.5 * 60 * 1000 # About 7 mins 30 secs
# endpos = 8.7 * 60 * 1000 # About 8 mins 40 secs
# cap.set(POS_MSEC, seekpos)
# Alternatively, we can search for a particular agent we're interested in:
# agent_to_search = 194
# ped_path = agents[agent_to_search]
# frame_num = timeframes[int(ped_path[0,0])]
# disp.set_frame(frame_num)
# disp.agent_num = agent_to_search
# disp.do_predictions = False
pl.ion()
display.plot_prediction_metrics([], [], [])
while cap.isOpened():
disp.do_frame()
key = cv2.waitKey(0) & 0xFF
if key == ord('e'):
run_experiment(cap, disp, timeframes, timesteps, agents)
elif key == ord('q') or key == ESC:
break
elif key == ord('r'):
disp.redo_prediction()
elif key == ord(' '):
disp.toggle_prediction()
elif key == LEFT:
disp.back_one_frame()
elif key == UP:
if disp.draw_all_agents: disp.next_sample()
else: disp.next_agent()
elif key == DOWN:
if disp.draw_all_agents: disp.prev_sample()
else: disp.prev_agent()
elif key == ord('a'):
disp.draw_all_agents = not disp.draw_all_agents
disp.draw_all_samples = not disp.draw_all_samples
disp.reset_frame()
elif key == ord('t'):
disp.draw_truth = not disp.draw_truth
disp.reset_frame()
elif key == ord('f'):
disp.draw_plan = not disp.draw_plan
disp.reset_frame()
elif key == ord('p'):
disp.draw_past = not disp.draw_past
disp.reset_frame()
elif key == ord('s'):
disp.draw_samples = (disp.draw_samples +
1) % display.SAMPLE_CHOICES
disp.reset_frame()
cap.release()
cv2.destroyAllWindows()
pl.close('all')
def run_experiment(cap, disp, timeframes, timesteps, agents):
# We're going to compare ourselves to the agents with the following IDs.
print 'Running experiment...'
util.reset_timer()
agents_to_test = range(319, 331)
# agents_to_test = [48, 49]
true_paths = []
planned_paths = []
one_step_paths = []
# IGP_scores = np.zeros((len(agents_to_test), 2))
# ped_scores = np.zeros((len(agents_to_test), 2))
display.plot_prediction_metrics([], [], [])
for i, agent in enumerate(agents_to_test):
print 'Agent ', agent
ped_path = agents[agent]
path_length = ped_path.shape[0]
start = 3 # Initializes path with start+1 points.
final_path = np.zeros((path_length, 3))
final_path[0:start + 1, :] = ped_path[0:start + 1, 1:4]
one_step_final_path = final_path.copy()
# Run IGP through the path sequence.
for t in range(start, path_length - 1):
frame_num = timeframes[int(ped_path[t, 0])]
print 'doing frame', frame_num
disp.set_frame(frame_num)
final_path[t + 1], one_step_final_path[t + 1] = disp.do_frame(
agent, final_path[:t + 1], False, True)
if cv2.waitKey(1) != -1:
print 'Canceled!'
return
# Compute the final score for both IGP and pedestrian ground truth.
print 'Agent', agent, 'done.'
# start_time = int(ped_path[0,0])
# other_peds = [agents[a] for a in timesteps[start_time] if a != agent]
# other_paths = [util.get_path_at_time(start_time, fullpath)[1][:,1:4] for fullpath in other_peds]
# IGP_scores[i] = util.length_and_safety(final_path, other_paths)
# ped_scores[i] = util.length_and_safety(ped_path[:,1:4], other_paths)
true_paths.append(ped_path[:, 1:4])
planned_paths.append(final_path)
one_step_paths.append(one_step_final_path)
pred_errs = util.calc_pred_scores(true_paths, planned_paths,
util.prediction_errors)
path_errs = util.calc_pred_scores(true_paths, planned_paths,
util.path_errors)
display.plot_prediction_metrics(pred_errs, path_errs,
agents_to_test[0:i + 1])
one_step_errors = util.calc_pred_scores(true_paths, one_step_paths,
util.prediction_errors)
pl.figure(2)
pl.clf()
display.plot_prediction_error('One-Step Prediction Errors',
one_step_errors, agents_to_test[0:i + 1])
pl.draw()
# results = np.column_stack((agents_to_test, ped_scores, IGP_scores))
pred_results = np.column_stack((agents_to_test, pred_errs.T))
path_results = np.column_stack((agents_to_test, path_errs.T))
# np.savetxt('experiment.txt', results)
np.savetxt('prediction_errors.txt', pred_results)
np.savetxt('path_errors.txt', path_results)
print 'EXPERIMENT COMPLETE.'
util.report_time()
def main():
# Parse command-line arguments.
args = parse_args()
Hfile = os.path.join(args.datadir, "H.txt")
mapfile = os.path.join(args.datadir, "map.png")
obsfile = os.path.join(args.datadir, "obsmat.txt")
destfile = os.path.join(args.datadir, "destinations.txt")
# Parse homography matrix.
H = np.loadtxt(Hfile)
Hinv = np.linalg.inv(H)
# Parse obstacle map.
obs_map = ewap.create_obstacle_map(mapfile)
# Parse pedestrian annotations.
frames, timeframes, timesteps, agents = ewap.parse_annotations(Hinv, obsfile)
# Parse destinations.
destinations = np.loadtxt(destfile)
# Play the video
seqname = os.path.basename(args.datadir)
cap = cv2.VideoCapture(os.path.join(args.datadir, seqname+".avi"))
disp = display.Display(cap, Hinv, obs_map, frames, timesteps, agents, destinations)
disp.model = models.model21
# We can set the exact frame we want to look at:
# disp.set_frame(780) # ETH sequence, beginning
# disp.set_frame(2238) # ETH sequence, agent #48 (1 v 1)
disp.set_frame(11301) # ETH sequence, big crowds both ways
# disp.set_frame(8289) # ETH sequence, agent #175 (1 v crowd)
# disp.set_frame(9867) # ETH sequence, agent #236 (1 v crowd)
# disp.set_frame(8859) # ETH sequence, agent #194 (1 v 1, choosing non-optimal avoidance)
# disp.set_frame(9261) # Hotel sequence, agent #175
# We can also go by minutes & seconds:
# seekpos = 7.5 * 60 * 1000 # About 7 mins 30 secs
# endpos = 8.7 * 60 * 1000 # About 8 mins 40 secs
# cap.set(POS_MSEC, seekpos)
# Alternatively, we can search for a particular agent we're interested in:
# agent_to_search = 194
# ped_path = agents[agent_to_search]
# frame_num = timeframes[int(ped_path[0,0])]
# disp.set_frame(frame_num)
# disp.agent_num = agent_to_search
# disp.do_predictions = False
pl.ion()
display.plot_prediction_metrics([], [], [])
while cap.isOpened():
disp.do_frame()
key = cv2.waitKey(0) & 0xFF
if key == ord('e'):
run_experiment(cap, disp, timeframes, timesteps, agents)
elif key == ord('q') or key == ESC:
break
elif key == ord('r'):
disp.redo_prediction()
elif key == ord(' '):
disp.toggle_prediction()
elif key == LEFT:
disp.back_one_frame()
elif key == UP:
if disp.draw_all_agents: disp.next_sample()
else: disp.next_agent()
elif key == DOWN:
if disp.draw_all_agents: disp.prev_sample()
else: disp.prev_agent()
elif key == ord('a'):
disp.draw_all_agents = not disp.draw_all_agents
disp.draw_all_samples = not disp.draw_all_samples
disp.reset_frame()
elif key == ord('t'):
disp.draw_truth = not disp.draw_truth
disp.reset_frame()
elif key == ord('f'):
disp.draw_plan = not disp.draw_plan
disp.reset_frame()
elif key == ord('p'):
disp.draw_past = not disp.draw_past
disp.reset_frame()
elif key == ord('s'):
disp.draw_samples = (disp.draw_samples + 1) % display.SAMPLE_CHOICES
disp.reset_frame()
cap.release()
cv2.destroyAllWindows()
pl.close('all')
