anomaly_detector

This repository contains the source code accompanying the Master Thesis of Ludwig Waffenschmidt with the title "Anomaly detection for vision-based obstacle detection in autonomous vehicles" presented at the Laboratory for Machine Tools and Production Engineering (WZL) at RWTH Aachen University.

The goal is the identification of obstacles in images using a deep hybrid anomaly detection approach in which a deep neural network extracts high level features which are then classified using a conventional classification algorithm.

Install

python -m virtualenv .env           # Create virtualenv
source .env/bin/                    # Activate it
pip install -r requirements.txt     # Install the python dependencies

pip install -e .    # Install current directory as editable pip package

Requires TensorFlow 2.1 for the feature extraction part.

Note that if you want to be able to use the rosbag_to_... scripts to extract images and metadata from bag files you need to have at least

• a bare bones ROS (Kinetic) and
• the cv_bridge package installed (sudo apt-get install ros-kinetic-cv-bridge).

Constants

Create a file ./anomaly_detector/consts.py with the following constants for quick debug excecutions:

# Paths
BASE_PATH      = "/path/to/base"
IMAGES_PATH    = BASE_PATH + "Images/"
EXTRACT_FILES  = IMAGES_PATH + "*.jpg"
EXTRACT_FILES_TEST = BASE_PATH + "Images/Test/*.jpg"
FEATURES_PATH  = BASE_PATH + "Features/"
BENCHMARK_PATH = BASE_PATH + "Benchmark/"
FEATURES_FILE  = FEATURES_PATH + "C3D.h5"
FEATURES_FILES = FEATURES_PATH + "*.h5"
METRICS_PATH   = FEATURES_PATH + "Metrics/"

# Defaults for feature extraction
DEFAULT_BATCH_SIZE = 128

Scripts

Bag file ➡ Images + Metadata ➡ Feature extractor ➡ Anomaly model

01_rosbag_to_images.py

Convert bag files to images with metadata files.

positional arguments:
  F                     The bag file(s) to convert. Supports "path/to/*.bag"

optional arguments:
  -h, --help            show this help message and exit
  --output_dir OUT      Output directory (default: {bag_file}/Images)
  --image_topic IM      Image topic (default: "/camera/color/image_raw")
  --image_crop X Y W H  Crop images using. Cropping is applied before scaling (default: Complete image)
  --image_scale SCALE   Scale images by this factor (default: 1.0)
  --tf_map TF_M         TF reference frame (default: map)
  --tf_base_link TF_B   TF camera frame (default: base_link)
  --override            Override existing images (default: False)
  --label L              0: Unknown (default)
                         1: No anomaly
                         2: Contains an anomaly

02_relabel.py

Label images.

optional arguments:
  -h, --help  show this help message and exit
  --images F  Path to images (default: /path/to/images/from/consts.py/)

03_extract_features.py

Extract features from images.

optional arguments:
  -h, --help            show this help message and exit
  --list                List all extractors and exit
  --files [F [F ...]]   File(s) to use (*.jpg)
  --extractor [EXT [EXT ...]]
                        Extractor name. Leave empty for all extractors (default: "")

04_rasterization_and_models_parallel.sh

Run this instead of 04x_rasterization_and_models.py to utilize multiple CPU cores

04x_rasterization_and_models.py

This script calculates

• patch locations
• rasterizations
• models + mahalanobis distances. These calculations - especially rasterization and mahalanobis distances - are very slow and only seem to use one core. You can call 04_rasterization_and_models_parallel.sh instead to run multiple instances of this script, which will then utilize more CPU cores. But beware heavy RAM use!

optional arguments:
  -h, --help           show this help message and exit
  --files [F [F ...]]  The feature file(s). Supports "path/to/*.h5"
  --index I            Used for parallelization (see 04_rasterization_and_models_parallel.sh)
  --total T            Used for parallelization (see 04_rasterization_and_models_parallel.sh)

05_metrics.py

Calculate metrics for the specified anomaly models.

optional arguments:
  -h, --help           show this help message and exit
  --files [F [F ...]]  The feature file(s). Supports "path/to/*.h5"
  --output OUT         Output file (default: "")

06_feature_extractor_benchmark.py

Benchmark the specified feature extractors.

optional arguments:
  -h, --help            show this help message and exit
  --files F             File(s) to use for benchmarks (*.tfrecord, *.jpg)
  --extractor [EXT [EXT ...]]
                        Extractor name. Leave empty for all extractors (default: "")
  --output OUT          Output file (default: "")
  --batch_sizes [B [B ...]]
                        Batch size for testing batched extraction. (default: [8,16,32,64,128,256,512])
  --init_repeat B       Number of initialization repetitions. (default: 3)
  --extract_single_repeat B
                        Number of single extraction repetitions. (default: 100)
  --extract_batch_repeat B
                        Number of batch extraction repetitions. (default: 10)

07_rasterization_benchmark.py

Benchmark rasterization for spatial binning.

optional arguments:
  -h, --help           show this help message and exit
  --files [F [F ...]]  The feature file(s). Supports "path/to/*.h5"
  --output OUT         Output file (default: "")

08_anomaly_model_benchmark.py

Benchmark the specified anomaly models.

optional arguments:
  -h, --help           show this help message and exit
  --files [F [F ...]]  The feature file(s). Supports "path/to/*.h5"
  --output OUT         Output file (default: "")