Algorithm 3: Trajectory Recovery.

Input: query feature $\mathrm{q}$ ; road network $\mathrm{N}$ ; feature gallery $\mathrm{F}$ ; threshold $\mathsf{\delta }$  Output: complete trajectory $\mathrm{P}$ 1 Load all the spatial-temporal knowledge with $\mathrm{N}$ ; 2 Search $\mathrm{k}$ of the most similar features in F and denote them as $\mathrm{top}\_\mathrm{k}$ ; 3 Initialize proximity graph $\mathrm{G}$ ; 4 for $\left({\mathrm{n}}_{\mathrm{i}},{\mathrm{n}}_{\mathrm{j}}\right)\in$ $\mathrm{top}\_\mathrm{k}$  do 5   Calculate coherence of $\left({\mathrm{n}}_{\mathrm{i}},{\mathrm{n}}_{\mathrm{j}}\right)$ and denote it as $s$ ; 6   if $\mathrm{s}$ > $\mathsf{\delta },$  then 7     $G.AddEdge\left({\mathrm{n}}_{\mathrm{i}}\to {\mathrm{n}}_{\mathrm{j}}\right)$ ; 8  ${\mathrm{P}}_{\mathrm{c}}\leftarrow \mathrm{CoarsePathSelection}\left(\mathrm{G}\right)$ ; 9  ${\mathrm{P}}_{\mathrm{f}}\leftarrow \mathrm{FinePathSelection}\left({\mathrm{P}}_{\mathrm{c}}\right)$ ; 10  ${\mathrm{P}}_{\mathrm{m}}\leftarrow \mathrm{Merge}\left({\mathrm{P}}_{\mathrm{f}}\right)$ ; 11  $\mathrm{P}\leftarrow \mathrm{Refine}\left({\mathrm{P}}_{\mathrm{m}}\right)$ ; 12 return $\mathrm{P}$ ;