Urban traffic velocity estimation method based on multi-source crowd sensing data
The present invention discloses an urban traffic velocity estimation method based on multi-source crowd sensing data. This method, based on roadside pedestrian data and road navigation data collected by smart phones, obtains a final estimated velocity through the steps of missing data filling, self-view velocity aggregation and multi-view velocity fusion. This fine-grained large-scale urban traffic velocity estimation method can achieve velocity estimation on all types of roads, including suburban road sections and paths, instead of just focusing on main roads in a city center. According to the present invention, based on data driving, the urban traffic velocity estimation method does not need to install additional devices on roads, and is low in cost and high in universality. Compared with the prior art, the urban traffic velocity estimation method has higher practicability, theoretical property and applicability, and is of great significance for improving traffic management and planning.
1 . An urban traffic velocity estimation method based on multi-source crowd sensing data, comprising the following steps:
step 1, data set preprocessing: filtering, by a processor, an original data set collected by smart phones to obtain roadside pedestrian data and road navigation data respectively;
step 2, average velocity calculation: calculating, by the processor, a current velocity X of each road section in different time periods by using the preprocessed data set in the step 1;
step 3, missing data filling: by using the current velocity X of each road section in different time periods obtained in the step 2, in combination with a historical velocity H of each road section in different time periods, filling, by the processor, missing data in X to obtain a filled velocity {circumflex over (X)}, reducing cost of velocity estimation, and does not require additional installation of devices;
wherein the step 3 specifically comprising:
firstly, introducing a mask matrix M to represent a missing unit of the velocity X:
M
i
,
j
=
{
1
,
if
X
i
,
j
is
available
0
,
if
X
i
,
j
is
missing
secondly, establishing a historical velocity matrix H by using historical data to provide additional information to help fill in missing data, and introducing another mask matrix N to represent a missing unit of the historical velocity matrix H:
N
i
,
j
=
{
1
,
if
H
i
,
j
is
available
0
,
if
H
i
,
j
is
missing
where i and j represent road section,
defining a weighted matrix W to measure an importance of each item in the historical velocity matrix H, and then performing matrix decomposition by using H, W, M, N and X:
Loss
h
(
W
)
=
1
2
W
⊙
N
⊙
(
H
-
UV
T
)
2
+
λ
U
2
+
λ
V
2
U
+
(
W
)
=
U
-
α
∂
Loss
h
(
W
)
∂
U
V
+
(
W
)
=
V
-
α
∂
Loss
h
(
W
)
∂
V
then constructing a loss by the matrix decomposition based on the updated U + and V + :
Loss
t
(
W
)
=
1
2
M
⊙
(
X
-
U
+
(
V
+
)
T
)
2
+
λ
U
+
2
+
λ
V
+
2
then updating a weight W:
W
+
=
W
-
β
∂
Loss
t
∂
W
where, α and β are learning rate parameters; and
after iterative updating, obtaining a learned weighted matrix W, X and H being able to be used simultaneously to estimate the missing data {circumflex over (X)}:
min
U
.
V
1
2
M
⊙
(
H
-
UV
T
)
2
+
1
2
W
⊙
N
⊙
(
H
-
UV
T
)
2
+
λ
U
2
+
λ
V
2
X
^
=
M
⊙
X
+
(
1
-
M
)
⊙
UV
T
where, λ represents a penalty item parameter, and U and V represent two sub-matrices decomposed from an original matrix;
step 4, self-view velocity aggregation: by using the filled velocity {circumflex over (X)} calculated in the step 3, quantizing, by the processor, spatial dependences between different road sections according to the historical velocity H, and collecting, by the processor, neighbor information to obtain an aggregated roadside pedestrian velocity data V d and road mobile navigation velocity data V w ; and
step 5, multi-view velocity fusion: according to the aggregated roadside pedestrian velocity data V d and the road mobile navigation velocity data V w , fusing, by the processor, multi-source velocity data by using a multi-layer perceptron according to a determination whether a time stamp and current velocity data are filled data, to obtain a fusion velocity Ŷ, and finally correcting, by the processor, the fusion velocity according to a feature of a road type to obtain a fine-grained large-scale urban traffic estimated velocity Ŷ′, wherein the road type includes road sections, main road, secondary road, tertiary road and other road.
2 . The urban traffic velocity estimation method based on multi-source crowd sensing data according to claim 1 , wherein the step 1 specifically comprises: obtaining the road navigation data by filtering a mobile application with location-based services usage list in original data; and by filtering a scanned wireless communication signal list in the original data, obtaining the roadside pedestrian data according to a determination whether there is a vehicle-mounted wireless communication signal in the list.
3 . The urban traffic velocity estimation method based on multi-source crowd sensing data according to claim 1 , wherein the step 2 specifically comprises: by using the data obtained by filtering in the step 1, projecting trajectory data into a road network by using a hidden Markov road network matching algorithm, so as to obtain the current velocity X of each road section in different time periods.
4 . The urban traffic velocity estimation method based on multi-source crowd sensing data according to claim 1 , wherein the step 4 specifically comprises: capturing a spatial correlation between adjacent roads by using the self-view velocity aggregation, and aggregating information of neighbor road sections correlated to a central road section; firstly, calculating a spatial correlation e i,j between a road section i and a road section j according to a historical velocity matrix, and keeping correlated parts and ignoring irrelevant information:
e
i
,
j
′
=
∑
(
H
i
,
:
-
H
j
,
:
)
2
,
j
∈
the
set
of
neighboring
connected
roads
for
i
e
i
,
j
=
{
+
∞
,
e
i
,
j
′
≥
threshold
e
i
,
j
′
,
e
i
,
j
′
≤
threshold
then, calculating a fusion coefficient a i,j between the road sections according to the spatial correlation e i,j , and then obtaining the roadside pedestrian velocity data V d and the road mobile navigation velocity data V w after a self-view aggregation:
a
i
,
j
=
exp
(
-
e
i
,
j
/
k
)
∑
j
=
N
i
(
-
e
i
,
j
k
)
+
ε
V
i
,
=
(
1
-
∑
j
=
N
i
a
i
,
j
2
)
*
X
^
i
,
:
+
∑
j
=
N
i
a
i
,
j
2
*
X
^
j
,
:
where, ε represents a minimal constant to prevent an overflow, k represents a constant scaling value, N i represents a set of neighboring connected road sections i, and V i , uniformly represents a feature representation of an i-th row, i.e., road section i of the roadside pedestrian velocity data V d and the road mobile navigation velocity data V w .
5 . The urban traffic velocity estimation method based on multi-source crowd sensing data according to claim 1 , wherein the step 5 specifically comprises: fusing the multi-source velocity data by using multi-view velocities, according to a determination whether a feature representing the time stamp is a filled data feature F d and whether the current velocity data is a filled data feature F w , then passing the features through an embedding layer and splicing the features, and according to the aggregated roadside pedestrian velocity data V d and road mobile navigation velocity data V w obtained in step 4, obtaining the fusion velocity Ŷ through the multi-layer perceptron:
Z
=
Embedding
(
Concat
(
F
d
,
F
w
)
)
Y
ˆ
=
MLP
(
Concat
(
Z
,
V
d
,
V
w
)
)
finally, correcting the estimated velocity according to the features of the road types, obtaining a velocity distribution of each type of road according to the historical data, regarding the velocity distribution as a normal distribution, and correcting the velocity falling at a tail of the distribution to obtain a final estimated velocity Ŷ′.