Multi-Camera Tracking for Yolov5 + ByteTrack, with S3 exporting support

Features:

This repo is a reorganized version for people tracking, based on Yolov5 + ByteTrack, features currently supported are the following:

1. multiple real-time streams / multple video files / local cameras inference support
2. detail export for each track (track_id, best_image in base64, large_image in base64, start_time, end_time, trajectory, camera_id in coordinates, etc.)
3. Closure of unfinished tracks support
4. Upload images, trajectory info to S3 support
5. Visualization enhanced - support trajectory visualization

config.yaml

config.yaml file notes the configurations for reid on cloud, including [TBD]

Inference Modes:

First, install requirements.

pip install -r requirements.txt

Single Source Test

Track.py does people tracking on an local file and gives output (mp4, json) in ./runs/exp[last epoch] and visualize bounding boxes and tracks. Start local inference on one file:

python track.py --local-mode  --show-vid  --source abc.mp4

Multiple Source Test

for multiple camera from video file inference, edit source json file, specify whether it is video_file or stream

python track.py --local-mode  --show-vid --save-vid --use-local-json-file --save-to-numpy-sample  --source ./local_test_2.json

Local Camera Test

for single local camera (mac cameras tested) inference

python track.py --local-mode --use-local-camera --show-vid --save-vid --save-to-numpy-sample

Sagemaker Notebook Server

Sagemaker Notebook Server mode is for deployment on sagemaker, and read & write files to S3. For further instruction, plz send us a ticket for more details. The follow should be run in order to start the server mode

python track.py

Send requests to process files on S3

TO-DOs

None