Switching IDs when tracking vehicles

Thank you for the repo!

Issues looked at #91 , #133, #76

Current Behavior

I am using fastmot for vehicle re-identification and tracking. There is a lot of id switching for stationary cars and cars traveling at low/high speeds. I am using SSD InceptionV2 as detector and VeRiWild as feature extractor. The input video is interesting as it contains all types of objects small, fast, slow, stationary, etc.

Input video: Input video Output video: output video Config file: config file

Adjusted parameters: max_reid_cost as reid is accurate, std_factor_acc, std_offset_acc from #76.

I was able to save output video with arguments -o path/to/output.mp4 but the output video contains black screen.

How to Reproduce

Added to reid.py

class VeRI(ReID):
    ENGINE_PATH = Path(__file__).parent / 'veriwild_r50ibn.trt'
    MODEL_PATH = Path(__file__).parent / 'veriwild_r50ibn.onnx'
    INPUT_SHAPE = (3, 256, 256)
    OUTPUT_LAYOUT = 2048
    METRIC = 'cosine'

Run the app.py with given config file.

Describe what you want to do

1. What input videos you will provide, if any: Input video
2. What outputs you are expecting: less switching
3. Ask your questions here, if any:

Your Environment

• Desktop
  • Operating System and version: Ubuntu 18.04
  • NVIDIA Driver version: 470.57.02
  • Used the docker image? Yes

Common issues

1. GStreamer warnings are normal
2. If you have issues with GStreamer on Desktop, disable GStreamer and build FFMPEG instead in Dockerfile
3. TensorRT plugin and engine files have to be built on the target platform and cannot be copied from a different architecture
4. Reinstalled OpenCV is usually not as optimized as the one shipped in Jetpack

References for models, labels, videos and config: GDrive link

[GeekAlexis]

These are not duplicate tracks. The issue is from the YOLO outputs. Duplicate detections can’t be removed. Because a new duplicate track will be created if YOLO keeps outputting duplicates. remove_duplicate() is deprecated in version 2 and replaced by rectify_tracks(), which is more robust in crowded scenes.

[GeekAlexis]

Yes that would help.