Railcar Tracking

Railcar Tracking is a framework to monitor railroad webcams from multiple sources. When moving train is detected, railcars are counted and classified and replay of passing train is saved.

Requirements

• ffmpeg
• youtube-dl
• Python>=3.8
• requirements.txt

Installing

Make sure you add ffmpeg and youtube-dl to system path. Don't forget to install requirements.txt.

Railcar Tracker

Railcar Tracker is a tool to count railcars and classify them.

Usage

track_example.py

from ultralytics import YOLO
from tracker.object_counter import count_objects

model_path = '../models/yolov8n_v2.pt'
model = YOLO(model_path)
video_path = '../video_examples/stat_1-1_720-30.mp4'

# see tracker_cfg.yaml for all kwargs
counter = count_objects(model, video_path, verbose=True)
print(counter, '\nTotal:', sum(counter.values()))

Detection and Tracking

Result

Counter({'boxcar': 48, 'tank': 29, 'hopper': 7, 'gondola': 5, 'locomotive': 3}) 
Total: 92

Tracker Config

You can use custom config for tracking, if it's inside config/:

counter = count_objects(model, video_path, config='custom_tracker_cfg.yaml')

Default config is tracker_cfg.yaml

# Settings for tracker and object detection
# You can also use parameters listed in https://docs.ultralytics.com/modes/predict/#inference-arguments

# General settings
device: None    # device to run on, i.e. cuda device=0/1/2/3 or device=cpu
save: False    # save video with results

# Object detection settings
conf: 0.6    # object confidence threshold for detection
iou: 0.4    # intersection over union (IoU) threshold for NMS
max_det: 300    # maximum number of detections per frame

# Tracker settings
tracker: custom_fast.yaml    # tracker config, see ultralytics/tracker/cfg
stream: True    # whether the input source is a video stream
persist: False    # whether to persist the trackers if they already exist
vid_stride: 1    # video frame-rate stride

# Debug settings
show: True    # stream output during inference
debug: False    # print counter debug info
verbose: True    # whether the inference is verbose or not

# Show settings
line_width: 2    # the line width of the bounding boxes, if None, it is scaled to the image size
show_labels: True    # show box labels
show_conf: True    # show confidence score
boxes: True    # show boxes in segmentation predictions

Clipper

Clipper is a tool to save short clips of YouTube and RTSP streams. Those clips are used to detect moving trains.

Usage

track_example.py

from recorder.clipper import yt_clip

yt_clip('https://www.youtube.com/watch?v=aWVfQ4xk_kU', '../replays/clips',
        ext='mp4', duration=5, fps=7, max_height=720)

Result

example.mp4, containing last 5 seconds of YouTube stream with 7 fps and 1280x720 size.

Dataset

TODO

Models

TODO