HaydenFaulkner / BicycleDetection

BicycleDetection

Install

Make two symbolically linked (shortcut) directories data and models within this project. Each will store the data and the models respectively.

cd path/to/BicycleDetection

# From within the BicycleDetection project directory
ln -s source/with/data data
ln -s source/with/models models

Data processing

Make another directory data/unfiltered and put the videos and .saa annotation files into this directory, if they are in subdirectories that is fine, however make sure all filenames are unique. Video files are matched with annotation files based on filenames. Files are filtered into a data/filtered directory using

Will need to use aligner to align .saa files

Training

You can skip this section

Use train_yolo.py in the run directory

Usage

In the main directory there are six python scripts available to run to perform different things. The main two are:

• full.py takes videos held in data/unprocessed and extracts frames, detects, tracks, and visualises. It will create a number of directories to store different data and visualisations:

  • data/stats to store video stats .txt files
  • data/frames to store extracted frames as .jpgs
  • data/detections to store detection .txt files
  • data/tracks to store tracking .txt files
  • data/snapshots to store image .jpg and video .mp4 snapshots of each track
  • data/summaries to store image .jpg and video .mp4 summaries for each video

• subclipper.py will take videos held in data/to_shorten and output shortened clips to data\shortened, performing frame extraction, detection, and tracking if necessary.

The other four can be separately or will be called by full.py and subclipper.py automatically:

• video_to_frames.py is used to extract frames from a video, the frames are used in the detector and are loaded more efficiently than getting them directly from the video;
• detect.py is used to detect cyclists in the images using a fast YOLOv3 + MobileNet1.0 object detection framework. This script will output a text file with the detections in the form: frame_num,class_id,confidence,left,top,right,bottom;
• track.py is used to associate the detections into tracks and will similarly output a text file with the tracks in the form: frame_num,track_id,confidence,left,top,right,bottom;
• visualise.py is used to visualise tracks and detections on the original video. It can also be used to snapshot individual cyclists as clips and/or cropped box images. It can also produce an overlay of all tracks in a video in a still;

You can check the available options for each script using the --help flag:

path/to/BicycleDetection$ full.py --help

You can change any parameter by using its flag and then the new value as so:

path/to/BicycleDetection$ full.py --display_tracks=false

path/to/BicycleDetection$ subclipper.py --model yolo --display_tracks

path/to/BicycleDetection$ subclipper.py --model frcnn --display_detections

Probably the two most important params for you is the --detect_every x (detect every x many frames) and the --model x (x is yolo or frcnn). The default is the yolo model.

The yolo model is the newer faster multi-gpu model that gives lower accuracies and might also still put the box higher than desired. The frcnn model is the old model which can only run on a single gpu, but is more accurate (it isn't too much slower than yolo), produces the 99% confidences. If you set --model frcnn you might also want to set --save_detection_threshold 0.99.