Detection of drones using their thermal signatures from thermal camera through YOLO-V3 based CNN with modifications to encapsulate drone motion

Last update: Dec 31, 2022

Overview

Drone Detection using Thermal Signature

This repository highlights the work for night-time drone detection using a using an Optris PI Lightweight thermal camera. The work is published in the International Conference of Unmanned Air Systems 2021 (ICUAS 2021) and the paper can be read in detail in ICUAS_2021_paper.

Requirements

The following are the requirements with Python 3.7.7

tensorflow==2.4.0
opencv_contrib_python==4.5.1.48
numpy==1.20.3

Model Architecture

The following diagram highlights the architecture of model based on YOLOV3. However, unlike typical single image object detection, the model takes in the concatenation of a specified number of images in the past relative to the image of interest. This is to encapsulate the motion of the drone as an input feature for detection, a necessity given that thermal signatures of different are generally globular in shape after a certain distance depending on the fidelity of the thermal camera used. Further details can be found in ICUAS_2021_paper.

Training and Testing

Clone the repository, adjust the training/testing parameters in train.py as shown and execute the code. The training data comprises of data from a controlled indoor environment while the test data contains a mixture data from indoor and outdoor environments.

# Train options
TRAIN_SAVE_BEST_ONLY        = True # saves only best model according validation loss (True recommended)
TRAIN_CLASSES               = "thermographic_data/classes.txt"
TRAIN_NUM_OF_CLASSES        = len(read_class_names(TRAIN_CLASSES))
TRAIN_MODEL_NAME            = "model_2"
TRAIN_ANNOT_PATH            = "thermographic_data/train" 
TRAIN_LOGDIR                = "log" + '/' + TRAIN_MODEL_NAME
TRAIN_CHECKPOINTS_FOLDER    = "checkpoints" + '/' + TRAIN_MODEL_NAME
TRAIN_BATCH_SIZE            = 4
TRAIN_INPUT_SIZE            = 416
TRAIN_FROM_CHECKPOINT       = False # "checkpoints/yolov3_custom"
TRAIN_LR_INIT               = 1e-4
TRAIN_LR_END                = 1e-6
TRAIN_WARMUP_EPOCHS         = 1
TRAIN_EPOCHS                = 10
TRAIN_DECAY                 = 0.8
TRAIN_DECAY_STEPS           = 50.0

# TEST options
TEST_ANNOT_PATH             = "thermographic_data/validate"
TEST_BATCH_SIZE             = 4
TEST_INPUT_SIZE             = 416
TEST_SCORE_THRESHOLD        = 0.3
TEST_IOU_THRESHOLD          = 0.45

Once the model is trained, you can test the model's predictions on images using detect_image.py. Adjust the the following parameters in detect_image.py and execute the code.

CLASSES               = "thermographic_data/classes.txt"
NUM_OF_CLASSES        = len(read_class_names(CLASSES))
MODEL_NAME            = "model_2"
CHECKPOINTS_FOLDER    = "checkpoints" + "/" + MODEL_NAME
ANNOT_PATH            = "thermographic_data/test/images/pr"
OUTPUT_PATH           = 'predicted_images/' + MODEL_NAME + "/pr"
DETECT_BATCH          = False
DETECT_WHOLE_VID      = True
BATCH_SIZE            = 1804
IMAGE_PATH            = ANNOT_PATH + "/free_3/free_3_frame_100"
INPUT_SIZE            = 416
SCORE_THRESHOLD       = 0.8
IOU_THRESHOLD         = 0.45

Similarly, you can test the model's predictions on videos using detect_video.py. Adjust the following parameters in detect_video.py and execute the code.

CLASSES               = "thermographic_data/classes.txt"
NUM_OF_CLASSES        = len(read_class_names(CLASSES))
MODEL_NAME            = "model_2"
CHECKPOINTS_FOLDER    = "checkpoints" + "/" + MODEL_NAME
ANNOT_PATH            = "raw_videos/free_2.mp4"
OUTPUT_PATH           = 'predicted_videos/' + MODEL_NAME 
INPUT_SIZE            = 416
SCORE_THRESHOLD       = 0.8
IOU_THRESHOLD         = 0.45

Examples of predictions

An example of correct drone detection in indoor environment shown below.

An example of correct drone detection in outdoor environment shown below.

Video of model predictions shown in indoor environment can be found here.

Detection of drones using their thermal signatures from thermal camera through YOLO-V3 based CNN with modifications to encapsulate drone motion

Related tags

Overview

Drone Detection using Thermal Signature

Requirements

Model Architecture

Training and Testing

Examples of predictions

Owner

Chong Yu Quan

Pytorch Implementation of Zero-Shot Image-to-Text Generation for Visual-Semantic Arithmetic

Regularizing Nighttime Weirdness: Efficient Self-supervised Monocular Depth Estimation in the Dark (ICCV 2021)

Offcial implementation of "A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Prediction, ICCV-2021".

A Decentralized Omnidirectional Visual-Inertial-UWB State Estimation System for Aerial Swar.

EDPN: Enhanced Deep Pyramid Network for Blurry Image Restoration

Styled Augmented Translation

Multi-task head pose estimation in-the-wild

A pure PyTorch implementation of the loss described in "Online Segment to Segment Neural Transduction"

A Comprehensive Empirical Study of Vision-Language Pre-trained Model for Supervised Cross-Modal Retrieval

Code for the Lovász-Softmax loss (CVPR 2018)

A simple python program that can be used to implement user authentication tokens into your program...

Code for sound field predictions in domains with impedance boundaries. Used for generating results from the paper

A small library for creating and manipulating custom JAX Pytree classes

You can draw the corresponding bounding box into the image and save it according to the result file (txt format) run by the tracker.

Conformer: Local Features Coupling Global Representations for Visual Recognition

RRL: Resnet as representation for Reinforcement Learning

ViSD4SA, a Vietnamese Span Detection for Aspect-based sentiment analysis dataset

[NeurIPS2021] Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks

On the Limits of Pseudo Ground Truth in Visual Camera Re-Localization

Rank 1st in the public leaderboard of ScanRefer (2021-03-18)