Sensing method and sensing device for human driven vehicles under partial VANET (vehicular ad hoc network) environment
A sensing method and a sensing device for HDVs (human driven vehicles) under a partial VANET environment are provided. According to the method, an existence sensing module and a location sensing module are constructed, based on a long-short-term-memory recurrent neural network, with utilizing historical information of motion states of a single CAV (connected and autonomous vehicle) as an input, existence and exact locations of surrounding HDVs of the CAV are outputted. The method is not only applicable to sensing the surrounding HDVs of the single CAV, but also the surrounding HDVs of the multiple CAVs. An estimation result of each CAV is firstly obtained, then the estimation results of the CAVs are checked with confliction criterion, according to checking results, the estimation results of the multiple CAVs are fused, and information about the surrounding HDVs of each CAV is finally outputted.
1. A sensing method for HDVs (human driven vehicles) under a partial VANET (vehicular ad hoc network) environment comprises steps of:
constructing an existence sensing module and a location sensing module, wherein: both of the existence sensing module and the location sensing module adopt state information of a CAV (connected and autonomous vehicle) as an input an output of the existence sensing module is existence of surrounding HDVs of the CAV; and an output of the location sensing module is probability density distributions of distances between the CAV and the surrounding HDVs;
for a scenario that a single CAV exists, adopting a first sensing procedure, specifically comprising steps of: estimating existence of each surrounding HDV of the CAV with the existence sensing module; estimating a location of each surrounding HDV with the location sensing module; continuously conducting the first sensing procedure, so as to output estimation results of the surrounding HDVs of the CAV in real-time;
for a scenario that multiple CAVs exist, adopting a second sensing procedure, specifically comprising steps of: applying the first sensing procedure to each CAV, so as to obtain existence and location information of surrounding HDVs of each CAV;
checking whether there is a confliction among the estimation results from the CAVs with confliction criterion; if the confliction exists, fusing conflicted estimation results, and obtaining final existence and location information of each HDV,
wherein: the existence sensing module is constructed based on a LSTM RNN (long-short-term-memory recurrent neural network), comprising two layers, wherein:
a first layer is a LSTM layer; an input of the first layer is historical motion information of the CAV, comprising lateral displacement, longitudinal displacement, normalized speed and acceleration; a sequence length of the first layer is set manually;
because state decision of the CAV needs to consider five vehicles, respectively an ego-lane leader, a left-lane leader, a left-lane follower, a right-lane leader, and a right lane follower of the CAV, an output of the first layer only adopts the last five data; and
a second layer is a sigmoid layer, for mapping the output of the first layer to an interval of [0,1]; and an output of the second layer is five scalars, respectively representing existence probabilities of the HDVs in five surrounding directions of the CAV.
2. A sensing method for HDVs (human driven vehicles) under a partial VANET (vehicular ad hoc network) environment comprises steps of:
constructing an existence sensing module and a location sensing module, wherein: both of the existence sensing module and the location sensing module adopt state information of a CAV (connected and autonomous vehicle) as an input an output of the existence sensing module is existence of surrounding HDVs of the CAV; and an output of the location sensing module is probability density distributions of distances between the CAV and the surrounding HDVs;
for a scenario that a single CAV exists, adopting a first sensing procedure, specifically comprising steps of: estimating existence of each surrounding HDV of the CAV with the existence sensing module; estimating a location of each surrounding HDV with the location sensing module; continuously conducting the first sensing procedure, so as to output estimation results of the surrounding HDVs of the CAV in real-time;
for a scenario that multiple CAVs exist, adopting a second sensing procedure, specifically comprising steps of: applying the first sensing procedure to each CAV, so as to obtain existence and location information of surrounding HDVs of each CAV; checking whether there is a confliction among the estimation results from the CAVs with confliction criterion; if the confliction exists, fusing conflicted estimation results, and obtaining final existence and location information of each HDV,
wherein: the location sensing module is constructed based on a LSTM RNN (long-short-term-memory recurrent neural network), comprising three layers, wherein:
a first layer is a LSTM layer; an input of the first layer is historical motion information of the CAV, comprising lateral displacement, longitudinal displacement, normalized speed and acceleration;
a second layer is a feed forward layer; the second layer adopts an output of the first layer as an input; and an output of the second layer is five classes of variables, respectively corresponding to five surrounding directions of the CAV; and
a third layer is a mixture density layer, which constructs a probability density distribution function of locations of the HDVs based on a GMM (Gaussian mixture model) with utilizing the output of the second layer, and obtains estimates.
3. A sensing method for HDVs (human driven vehicles) under a partial VANET (vehicular ad hoc network) environment comprises steps of:
constructing an existence sensing module and a location sensing module, wherein: both of the existence sensing module and the location sensing module adopt state information of a CAV (connected and autonomous vehicle) as an input an output of the existence sensing module is existence of surrounding HDVs of the CAV; and an output of the location sensing module is probability density distributions of distances between the CAV and the surrounding HDVs;
for a scenario that a single CAV exists, adopting a first sensing procedure, specifically comprising steps of: estimating existence of each surrounding HDV of the CAV with the existence sensing module; estimating a location of each surrounding HDV with the location sensing module; continuously conducting the first sensing procedure, so as to output estimation results of the surrounding HDVs of the CAV in real-time;
for a scenario that multiple CAVs exist, adopting a second sensing procedure, specifically comprising steps of: applying the first sensing procedure to each CAV, so as to obtain existence and location information of surrounding HDVs of each CAV; checking whether there is a confliction among the estimation results from the CAVs with confliction criterion; if the confliction exists, fusing conflicted estimation results, and obtaining final existence and location information of each HDV,
wherein the confliction criterion is: if two CAVs simultaneously satisfy following three rules, a confliction exists between estimation results of the two CAVs, wherein:
a first rule is that the two CAVS are separated by at most one lane;
a second rule is that a longitudinal distance between the two CAVs is smaller than a distance threshold; and
a third rule is that the two CAVs and surrounding HDVs thereof are probabilistic close; a meaning of probabilistic close is defined as follows:
{
∫
mi
n
(
x
^
M
EV
1
·
x
^
N
EV
2
)
ma
x
(
x
^
M
EV
1
·
x
^
N
EV
2
)
p
M
EV1
(
x
^
M
EV
1
-
x
)
dx
≤
p
=
∫
mi
n
(
x
^
M
EV
1
·
x
^
N
EV
2
)
ma
x
(
x
^
M
EV
1
·
x
^
N
EV
2
)
p
N
EV
2
(
x
^
N
EV
2
-
x
)
dx
≤
p
=
wherein: EV 1 and EV 2 respectively represent the two CAVs; M EV1 represents any one of surrounding HDVs of EV 1 ; N EV2 represents any one of surrounding HDVs of EV 2 ;
x
^
M
EV
1
and
x
^
N
EV
2
respectively represent location estimates of M EV1 and N EV2 ;
p
M
EV
1
(
x
)
and
p
N
EV
1
(
x
)
respectively represent probability density distributions of M EV1 and N EV2 ; and p is a set threshold.
4. The sensing method, as recited in claim 3 , wherein: when estimation results of N CAVs are conflicted, one result with maximum probability is selected from 1+N+1 possible results as a final estimation result; the 1+N+1 possible results are respectively:
(1) maintaining all of the estimation results of the N CAVs;
(2) maintaining an estimation result of one CAV, and discarding estimation results of the other CAVs, totally N possibility; and
(3) weighting probability densities in the estimation results of the N CAVs, and obtaining a final probability density and location estimate of each HDV.