Deep Alignment Network Based Multi-person Tracking with Occlusion and Motion Reasoning

• Introduction: This repository contains the implementation of paper Deep Alignment Network Based Multi-person Tracking with Occlusion and Motion Reasoning.

• Abstract: Tracking-by-detection is one of the most common paradigms for multi-person tracking, due to the availability of automatic pedestrian detectors. However, existing multi-person trackers are greatly challenged by misalignment in the pedestrian detectors (i.e., excessive background and part missing) and occlusion. To address these problems, we propose a deep alignment network based multi-person tracking method with occlusion and motion reasoning. Specifically, the inaccurate detections are firstly corrected via a deep alignment network, in which an alignment estimation module is used to automatically learn the spatial transformation of these detections. As a result, the deep features from our alignment network will have a better representation power and thus lead to more consistent tracks. Then, a coarse-to-fine schema is designed for construing a discriminative association cost matrix with spatial, motion and appearance information. Meanwhile, a principled approach is developed to allow our method to handle occlusion with motion reasoning and the re-identification ability of the pedestrian alignment network. Finally, a simple yet real-time Hungarian algorithm is employed to solve the association problem. Comprehensive experiments on MOT16, ISSIA soccer, PETS09 and TUD datasets validate the effectiveness and robustness of the proposed method.

Citation:

@article{zhou2018deep,
  title={Deep Alignment Network Based Multi-person Tracking with Occlusion and Motion Reasoning},
  author={Zhou, Qinqin and Zhong, Bineng and Zhang, Yulun and Li, Jun and Fu, Yun},
  journal={IEEE Transactions on Multimedia},
  year={2018},
  publisher={IEEE}
}

Preparation:

You need to compile the implementation of the Hungarian algorithm by running make.m in the tracking directory.

Usage:

1. Download the MOT16 sequences from https://motchallenge.net/data/MOT16/ and place them in data folder. Download the traied deep align model from https://pan.baidu.com/s/1UL5EfgvQJSRDbT3JIrMzXw (umdx) and place it in models folder.

2. Preparing Matconvnet and run 'gpu_compile.m' to compile the files used in establishing the deep appearance model.

3. You can run the other detector to obtain the detection results first. Simply put the corresponding object_02 folder underneath the corresponding sequence folder.

   • OR -

   Alternatively, you can also use the detections provided by https://motchallenge.net in the det folder.

4. To run the tracking stage, open 'tracker.m' and modify the variables base_dir and seq_dir to point to one of the downloaded sequences. Run the script. The tracking results are stored in 'tracking_results.txt'.

Experiments

1. Settings

• We conduct our experiments on 1 GTX1080ti GPU

2. Quantitative comparison results on the test sequences from the MOT16 benchmark

Tracker	MOTA	MOTP	MT	ML	FP	FN	ID SW.	Frag	Hz	Detector
JPDA_m	26.2	76.3	4.1%	67.5%	3689	130549	365	638	22.2	Public
GMPHD_HDA	30.5	75.4	4.6%	59.7%	5169	120970	539	731	13.6	Public
CppSORT	31.5	77.3	4.3%	59.9%	3048	120278	1587	2239	687.1	Public
CEM	33.2	75.8	7.8%	54.4%	6837	114322	642	731	0.3	Public
GM_PHD_N1T	33.3	76.8	5.5%	56.0%	1750	116452	3499	3594	9.9	Public
TBD	33.7	76.5	7.2%	54.2%	5804	112587	2418	2252	1.3	Public
HISP_T	35.9	76.1	7.8%	50.1%	6412	107918	2594	2298	4.8	Public
JCmin_MOT	36.7	75.9	7.5%	54.4%	2936	111890	667	831	14.8	Public
LTTSC-CRF	37.6	75.9	9.6%	55.2%	11969	101343	481	1012	0.6	Public
GMMCP	38.1	75.8	8.6%	50.9%	6607	105315	937	1669	0.5	Public
OVBT	38.4	75.4	7.5%	47.3%	11517	99463	1321	2140	0.3	Public
EAMTT_pub	38.8	75.1	7.9%	49.1%	8114	102452	965	1657	11.8	Public
Ours	40.8	74.4	13.7%	38.3%	15143	91792	1051	2210	6.5	Public

3. Quantitative comparison results on the PETS09-S2L2 benchmark

Tracker	MOTA	MOTP	MT	ML	FP	FN	ID Sw.	Frag	Detector
GMPHD_HDA	31.9	69.1	0.0%	31.0%	467	5965	131	315	Public
CppSORT	36.8	69.3	0.0%	16.7%	511	5251	331	480	Public
JPDA_m	37.6	65.9	11.9%	19.0%	1016	4858	139	260	Public
EAMTT_pub	39.9	69.7	7.1%	14.3%	758	4814	218	357	Public
Ours	44.7	68.6	9.5%	7.1%	1331	3707	294	546	Public

4. Quantitative comparison results on the TUD-Crossing benchmark

Tracker	MOTA	MOTP	MT	ML	FP	FN	ID Sw.	Frag	Detector
EAMTT_pub	48.0	72.9	23.1%	15.4%	110	436	27	37	Public
GMPHD_HDA	50.5	72.3	15.4%	15.4%	41	485	19	29	Public
CppSORT	50.6	74.0	7.7%	15.4%	22	489	33	57	Public
JPDA_m	60.9	68.4	30.8%	23.1%	44	385	2	26	Public
Ours	77.1	72.1	76.9%	0.0%	83	151	18	27	Public

5. Quantitative comparison results on the ISSIA soccer dataset

Tracker	MOTA	MOTP	MT	ML	FP	FN	ID Sw.	Detector
Ours with HSV	69.5	71.2	18.0%	5.0%	1463	2386	369	YOLOv3
Ours	77.1	77.9	19.0%	5.0%	1374	1992	119	YOLOv3

6. Quantitative tracking results obtained by our multi-person tracker using Faster R-CNN on MOT16 benchmark

Tracker	MOTA	MOTP	MT	ML	FP	FN	ID Sw.	Frag
Ours(with Faster R-CNN)	59.7	78.9	32.4%	21.6%	11034	61160	1292	1575

Contacts:

If you have any question, please feel free to contact with me.

E-mail: [email protected]