FAANet: feature-aligned attention network for real-time multiple object tracking in UAV videos

Zhenqi Liang; Jingshi Wang; Gang Xiao; Liu Zeng

doi:10.3788/COL202220.081101

Journals >Chinese Optics Letters >Volume 20 >Issue 8 >Page 081101 > Article

Chinese Optics Letters
Vol. 20, Issue 8, 081101 (2022)

FAANet: feature-aligned attention network for real-time multiple object tracking in UAV videos

Zhenqi Liang¹, Jingshi Wang^1、2, Gang Xiao^1、*, and Liu Zeng¹

Author Affiliations

¹School of Aeronautics and Astronautics, Shanghai Jiao Tong University, Shanghai 200240, China

²Jiangsu Automation Research Institute, Lianyungang 222061, China

show less

DOI: 10.3788/COL202220.081101 Cite this Article Set citation alerts

Zhenqi Liang, Jingshi Wang, Gang Xiao, Liu Zeng. FAANet: feature-aligned attention network for real-time multiple object tracking in UAV videos[J]. Chinese Optics Letters, 2022, 20(8): 081101 Copy Citation Text

EndNote(RIS)

BibTex

Plain Text

show less

Architecture of our tracker FAANet tracking framework. This framework contains four components: backbone (RepVGG), neck (CSA + FAA), head (Re-ID + detection), and association.

Fig. 1. Architecture of our tracker FAANet tracking framework. This framework contains four components: backbone (RepVGG), neck (CSA + FAA), head (Re-ID + detection), and association.

Download full size | View in the Article

Fig. 2. Architecture of CSA module.

Download full size | View in the Article

Fig. 3. Architecture of FAA module.

Download full size | View in the Article

Fig. 4. Procedure of association between detections and tracklets.

Download full size | View in the Article

Fig. 5. Structural re-parameterization of a RepVGG block.

Download full size | View in the Article

Fig. 6. MOTA-IDF1-FPS comparison with other UAV-based MOT trackers on the UAVDT test dataset. The horizontal axis is FPS, the vertical axis is MOTA, and the radius of the circle is IDF1.

Download full size | View in the Article

Fig. 7. IDF1 comparison with other UAV-based MOT trackers on the UAVDT test dataset based on scene attributes. The IDF1 of FAANet is marked outside the circle.

Download full size | View in the Article

Fig. 8. Examples and comparison of tracking results between DeepSORT and FAANet on the UAVDT test dataset.

Download full size | View in the Article

MOT Methods	Year	Framework	MOTA $↑$	IDF1 $↑$	MOTP $↑$	MT $↑$	ML $↓$	FP $↓$	FN $↓$	IDS $↓$	FM $↓$	FPS $↑$
SORT^[1]	2016	Faster RCNN	39.0	43.7	74.3	33.9	28.0	33,037	172,628	2350	5787	Nan
DeepSORT^[2]	2017	Faster RCNN	40.7	58.2	73.2	41.7	23.7	44,868	155,290	2061	6432	15.01
DeepAlign^[20]	2018	Faster RCNN	41.6	49.0	73.3	43.7	24.3	45,420	152,224	1546	3733	0.23
SBMA^[21]	2019	LSTM	38.6	48.5	72.1	38.9	24.4	44,724	160,950	3489	11,796	Nan
IPGAT^[8]	2020	LSTM + CGAN	39.0	49.4	72.2	37.4	25.2	42,135	163,837	2091	10,057	Nan
M-CMSN-M^[9]	2020	Faster RCNN	43.1	62.6	73.5	45.3	22.7	45,900	147,638	390	4259	0.64
Quadruplet^[22]	2021	Faster RCNN	40.3	55.0	74.0	Nan	Nan	30,065	150,837	1091	3057	Nan
FAANet	Nan	RepVGG + JDE	44.0	64.6	77.9	47.9	22.6	57,146	133,496	403	7202	38.24

Table 1. Results of a Quantitative Comparison among Classic MOT Methods and Recent UAV-Based Methods on the UAVDT Test Dataset^a

View in the Article

RepVGG-B0	CA	SA	FAA	MOTA $↑$	IDF1 $↑$	FPS $↑$
$\sqrt$				38.2	56.8	45.70
$\sqrt$	$\sqrt$			39.7	59.2	43.52
$\sqrt$		$\sqrt$		39.3	59.4	43.41
$\sqrt$	$\sqrt$	$\sqrt$		40.4	60.2	41.35
$\sqrt$			$\sqrt$	42.1	63.7	40.54
$\sqrt$	$\sqrt$	$\sqrt$	$\sqrt$	44.0	64.6	38.24

Table 2. Evaluation of the Critical Factors in FAANet^a

View in the Article

Rep	Params (10⁶)	FLOPs (10⁹)	MOTA $↑$	IDF1 $↑$	FPS $↑$
	15.9	62.3	44.0	64.6	30.32
$\sqrt$	14.4	58.3	44.0	64.6	38.24

Table 3. The Improvement of Re-parameterization Technique

Zhenqi Liang, Jingshi Wang, Gang Xiao, Liu Zeng. FAANet: feature-aligned attention network for real-time multiple object tracking in UAV videos[J]. Chinese Optics Letters, 2022, 20(8): 081101

Download Citation

EndNote(RIS)

BibTex

Plain Text

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information