多目标跟踪-基于anchor-free目标检测+ReID的实时一阶多类多目标跟踪算法实现-优质项目实战.zip

# MCMOT: One-shot multi-class multi-object tracking </br> 单阶段实时多类别多目标跟踪 </br> This is an extention work of FairMOT, which extends the one-class multi-object tracking to multi-class multi-object tracking </br> You can refer to origin fork [FairMOT](https://github.com/ifzhang/FairMOT) ## Tracking demo of C5(car, bicycle, person, cyclist, tricycle)      </br> ## Tracking demo of VISDrone dataset       </br> ## VisDrone dataset training with 10 object classes </br> VisDrone is a public dataset for 4 CV challenges: object detection, crowd counting, single class multi-object tracking, multi-class multi-object tracking. * Download multi-class multi-object tracking part of Visdrone dataset * Using gen_dataset_visdrone.py script to generate labels. * Call the gen_dot_train_file function in gen_dataset_visdrone.py to generate the dot train file for VisDrone mcmot training task. * Uncomment cls2id and id2cls in multitracker.py to use the correct class names and class Ids mapping. ``` from gen_dataset_visdrone import cls2id, id2cls # visdrone # from gen_labels_detrac_mcmot import cls2id, id2cls # mcmot_c5 ``` * Set cls ids for visdrone training n opts.py i.e. ``` 1~10 object classes are what we need non-interest-zone (0) pedestrian (1) --> 0 people (2) --> 1 bicycle (3) --> 2 car (4) --> 3 van (5) --> 4 truck (6) --> 5 tricycle (7) --> 6 awning-tricycle (8) --> 7 bus (9) --> 8 motor (10) --> 9 others (11) self.parser.add_argument('--reid_cls_ids', default='0,1,2,3,4,5,6,7,8,9', # '0,1,2,3,4' or '0,1,2,3,4,5,6,7,8,9' help='') # the object classes need to do reid ``` ## Tracking or detection mode setting Set id_weight to 1 for tracking and 0 for detection mode. ``` self.parser.add_argument('--id_weight', type=float, default=1, # 0for detection only and 1 for detection and re-ida help='loss weight for id') # ReID feature extraction or not ``` ## Pretained model for C5 and VisDrone detection and tracking HRNet18 backbone with bi-linear upsampling replaced with de-convolution </br> The pre-trained model is for 5 classes(C5) detection & tracking: car, bicycle, person, cyclist, tricycle, which can be used for road traffic video surveillance and analysis. </br> </br> [baidu drive link](https://pan.baidu.com/s/1imrim0kt72_Ay9w-X4kqdw) extract code：ej4p </br> [one drive link](https://1drv.ms/u/s!AmgjV-Jny9CqkX2EneIkb6Q8x6WI?e=Lsqx4G) ### Resnet18 backbone for C5，which is much smaller than HRNet18 [ResNet18 one drive link](https://1drv.ms/u/s!AmgjV-Jny9Cqkjiwqx7UY2g2_uZU?e=tws0eQ) ### Resnet18 backbone for VisDrone mcmot [Resnet18 one drive link](https://1drv.ms/u/s!AmgjV-Jny9CqkjkCXS9QC8mJwMmB?e=XzrH1T) ### Using YOLOV4 as detector You can also refer to the ropo：[MCMOT_YOLOV4](https://github.com/CaptainEven/YOLOV4_MCMOT) </br> This is MCMOT with CenterNet detection frame work replaced with an anchor-based detection framework. ### Using ByteTrack You can also refer to the ropo：[MCMOT-ByteTrack](https://github.com/CaptainEven/MCMOT-ByteTrack) </br> Using YOLOX as front-end and using ByteTrack as back-end.