Symbolic representation and retrieval of moving object trajectories
read more
Citations
Clustering of Vehicle Trajectories
Visually mining and monitoring massive time series
Similarity search for multidimensional data sequences = 다차원 데이터 시퀀스에 대한 유사성 검색
Visualizing and discovering non-trivial patterns in large time series databases
Movement similarity assessment using symbolic representation of trajectories
References
Binary codes capable of correcting deletions, insertions, and reversals
Algorithms on Strings, Trees and Sequences: Computer Science and Computational Biology
Related Papers (5)
Frequently Asked Questions (19)
Q2. What have the authors stated for future works in "Symbolic representation and retrieval of moving object trajectories" ?
Future work includes the following problems: 1. Finding an embedding method, which keeps both the lower bound property and the temporal order of elements in the strings.
Q3. What is the pruning power of MPS?
MPS has quite stable pruning power over trajectory length, because it maintains in the strings, the order of the corresponding (movement direction, distance ratio) pairs, and its ability to remove a lot of false candidates due to its consideration of neighbors of each symbol.
Q4. How did they decompose the raw object sequences into components?
Chen and Chang [4] used wavelet transform to decompose raw object trajectories (position sequences) into components at different scale.
Q5. What is the simplest way to use MPS as a filter?
Using MPS as filter is based on the assumption that the retrieval cost may be reduced due to the smaller size of MPS compared to movement sequences.
Q6. how many neighbors of each integer point in the frequency space is there?
as the number of neighbors of each integer point in the frequency space is limited (at most 8), the computation time of Algorithm 3 is still linear.
Q7. How do the authors use frequency vectors to reduce the cost of computing NED?
the authors define NMFD between two frequency vectors and use frequency vectors as filters to save the cost of CPU time on computing NED.
Q8. How do the authors get the results for LCSS and NED?
The authors find that for ASL data, the authors get best results for LCSS and NED when ²dir = 0.167π and ²dis = 0.1 ∗ σmax, where σmax is the maximum value of movement distance ratio in the data set, which can be obtained when the authors convert raw trajectories to movement sequences.
Q9. what is the algorithm for quantizing a movement direction, distance ratio?
Once the authors quantize the (movement direction, distance ratio) space into subregions and derive the movement alphabet A, the authors use Algorithm 1 to map a (movement direction, distance ratio) pair (θ, σ) into a symbol.
Q10. What is the similarity measure that the authors propose?
The similarity measure that the authors propose takes the longest common subsequences, gap penalties and compared sequence lengths into consideration.
Q11. What is the algorithm for mapping a movement pattern string into a symbol?
Given a movement sequence MA = [(θa,1, σa,1), . . . , (θa,n, σa,n)] of length n and movement pattern alphabet A, a movement pattern string (MPS) is defined as a sequence of symbols: Sa,1Sa,2 . . .
Q12. What is the way to reduce false candidates in the retrieval of trajectory data?
Their experimental results confirm that NED is a suitable and superior similarity measure for trajectory data and feature vector with NMFD can effectively reduce the false candidates in trajectory retrieval.
Q13. What is the NED between the original movement sequences MA and MB?
NED between original movement sequences MA and MB is 0, whereas the NED between MPSA and MPSB that is computed based on the standard edit distance [20] is 1, which is not the lower bound of 0.
Q14. Why is the quantization map used to convert a (movement direction, distance ratio)?
This is because the (movement direction, distance ratio) pairs that are located near the boundary of quantization subregions may be assigned different symbols and require a replace operation that is not needed in the original sequence comparison.
Q15. Why does NED achieve the same number of correct results as LCSS?
Due to the lower bound property of NED on MPS, clustering on it achieves nearly the same number of correct results as that of clustering on original movement sequences.
Q16. How did they measure the distance between two trajectories?
Little and Gu [22] used the path and speed curves to represent the motion trajectories and measured the distance between two trajectories using DTW.
Q17. What is the difference between MPS and FV?
In terms of total retrieval efficiency, FV is much better than MPS due to the linearity of the computation cost of FV as opposed to quadratic cost for MPS.3.
Q18. What is the cost of converting movement patterns into MPS?
Even though the authors reduce the storage requirements by converting movement sequences into movement pattern strings, the cost of computing the NED between two MPSs is still O(n∗m), since the length of a movement sequence and that of its corresponding movement pattern string are the same.
Q19. what is the frequency distance between u and v?
Let u and v be integer points in s dimensional space, The frequency distance FD(u, v) between u and v is defined as the minimum number of steps that is required to go from u to v (or equivalently from v to u) by moving to a neighbor point at each step.