Occluded Video-Based Person Re-Identification Based on Spatial-Temporal Trajectory Fusion

Fig. 8. Parameter analysis experimental results. (a) Influence of temporal fusion loss value $T$ on Rank-1 and mAP; (b) influence of spatial fusion loss value $N$ on Rank-1 and mAP

Download full size

Fig. 9. Visualization of video-based person re-identification results

Download full size

Table 0. [in Chinese]

View table

Table 0. [in Chinese]

Algorithm 1：video-based person re-identification based on spatial-temporal trajectory fusion

Input： MARS_traj dataset；trajectory prediction model Social- GAN；video-based person re-identification model

Output： mAP and Rank-k

1）spatial coordinates and temporal information of person ID from query dataset of video sequences are input into Social-GAN model；

2）possible predicted trajectories are generated by generator in Social-GAN based on spatial coordinates and temporal information；

3）discriminator in Social-GAN discriminates generated predicted trajectory，and obtains matching predicted trajectory dataset query_pred.

4）For $i = 1 : N_{1}$ do

5） For $j = 1 : N_{2}$ do

6） temporal fusion loss $l_{j}^{t e m}$ and spatial fusion loss $l_{j}^{s a p}$ between $j$ -th video sequence in gallery and $i$ -th video prediction trajectory in query_pred are computed by equation（3）and（4），respectively；

7） end

8） $l_{i - j} = m i n (l_{j}^{t e m} + l_{j}^{s a p}); \forall j \in [1, N_{2}];$

9） value of $j$ corresponding to $l_{i - j}$ is obtained and assigned to $i_{j}$ ；

10） sent $i_{j}$ -th video sequence of gallery into query_TP；

11）end

12）fusion feature of query_TP and gallery are extracted，respectively；

13）feature distance metrics based on query_TP and gallery are calculated，and feature vectors of all gallery video sequences are ranked according to distance metrics；

14）probability of correct match within ranked gallery is calculated according to query；

15）return mAP and Rank-k.

Table 1. MARS_traj dataset
View table
Table 1. MARS_traj dataset
Subset MARS_traj
ID Tracklets
query 90 90
gallery 119 1271
total 119 1361

Table 2. Performance evaluation on MARS and MARS_traj datasets

View table

Table 2. Performance evaluation on MARS and MARS_traj datasets

Method	MARS		MARS_traj
Method	mAP	Rank-1	mAP	Rank-1
COSAM	80.50	81.20	71.90	69.40
COSAM+STTF			93.00	92.90
STE-NAVE	77.80	85.05	66.15	70.59
STE-NAVE+STTF			92.88	96.47
AP3D	84.10	89.10	62.90	62.40
AP3D+STTF			90.10	91.70
TCLNet	85.10	89.80	69.45	72.94
TCLNet+STTF			94.82	96.47

Tools

Get Citation

Copy Citation Text

Xiao Yun, Kaili Song, Xiaoguang Zhang, Xinchao Yuan. Occluded Video-Based Person Re-Identification Based on Spatial-Temporal Trajectory Fusion[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1010021

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites