Code base: https://github.com/xinshuoweng/AB3DMOT

Setting up environment:
Ubuntu 16.04, Python 3.7.6 in anaconda, python 1.2 + CUDA 9.2
refer to pointrcnn environment setting up and running for the pre research environment

Kitti object tracking dataset Download

Dataset official website: http://www.cvlibs.net/datasets/kitti/eval_Tracking. PHP
when pre researching according to the project description, you only need to download the image_ 02 dataset:
http://www.cvlibs.net/download.php?file=data_tracking_image_2.zip

Xinshuo_Pytoolbox preparation

The project requires the author’s toolkit https://github.com/xinshuoweng/Xinshuo_ Pytoolbox
you need to install the package described in its requirements.txt, and then put all the folders directly into the ab3dmot folder

step

  1. install dependeny：
$ cd path/to/AB3DMOT
$ pip install -r requirements.txt

  2. Inference - run our tracker on the KITTI MOT validation set with the provided detection:
$ python main.py pointrcnn_Car_val
$ python main.py pointrcnn_Pedestrian_val
$ python main.py pointrcnn_Cyclist_val
Just select one, e.g. pedestrian

  3. Extract the files in the image_2 package to data/KITTI/resources

  4. Change the variable seq_list in visualization.py to the sample number you want to visualize, e.g.
seq_list = ['0000', '0001', '0002', '0003', '0004', '0005', '0006', '0007', '0008'] 

  5. execute python visualization.py pointrcnn_Pedestrian_val (corresponds to pointrcnn_Pedestrian_val in step 2)

  6. Check the results in results/pointrcnn_Pedestrian_val/trk_image_vis

Effect evaluation

It is easy to recognize cuboids/columns as vehicles/pedestrians, especially when recognizing vehicles, but this misjudgment generally occurs outside the road area, – and the number of continuous frames is very small; It is possible to distinguish cyclists as ordinary pedestrians; There is no obvious missing judgment, but the tracking is likely to be interrupted when occlusion occurs, resulting in unstable tracking object ID;