ConvNet and LSH-Based Visual Localization Using Localized Sequence Matching

. 2019 May 28;19(11):2439. doi: 10.3390/s19112439

Algorithm 2 Localized sequence matching and visual localization.

Inputs:

{D}

{Cosine distance};

N^{t r a i n}, N^{t e s t}

{training and testing images numbers};

V_{m a x}, V_{m i n}

{maximum and minimum speed ratios};

V_{s t e p}

{Vehicle speed step-size}

d_{s}

{Sequence length};
Outputs:
S {Path-line (sequence candidate) score};
for T ←

d_{s}

N^{t e s t}

M^{T}

←

[D^{T - d_{s} + 1}, D^{T - d_{s} + 2}, \dots, D^{T}]

; // Local searching matrix.
j ← 1; // Path number (sequence candidates number) initialization.
for s ← 0 to

(N^{t r a i n} - V_{m a x} \times d_{s})

do
for V ←

V_{m i n}

V_{s t e p}

V_{m a x}

S_{T, j}

← 0;
for t ←

(T - d_{s} + 1)

to T do

k (t) \leftarrow s + V (t - (T - d_{s} + 1))

; // k is a line index in the column vector

D^{t}

; s is the training image number where the path originated in.

S_{T, j}

←

S_{T, j} + D_{k (t)}^{t}

; // Score S is calculated for each possible path.
end for
j ←

j + 1

; Sequence candidate number updating.
end for
end for

S S_{T} = \frac{min_{j} {S_{T, j}}}{min_{j (j \neq m 1)} {S_{T, j}}}

;

m 1

is the index of minimum score.
if

S S_{T} < = T h

Matching validation is positive;
Vehicle position ← The matched training image position
if

S S_{T} > T h

Matching validation is negative;
Vehicle position ← NaN (no position results)
end for