In recent decades, environmental crime has surged, posing threats to various environmental objectives [1]. Environmental damage is not always immediately visible. While some forms, like oil slicks, are evident, others, such as air pollution, remain concealed. The illicit trade of chemicals, particularly in the refrigeration sector, has contributed significantly to this rise, constituting 7.8% of global greenhouse gas emissions as measured in 2014 [2]. The Kyoto Protocol, an extension of the 1992 United Nations Framework Convention on Climate Change, establishes targets for reducing greenhouse gases, including hydrofluorocarbons (HFCs) [3].
While vital, restrictions on the production and use of ozone-depleting substances (ODS) inadvertently created a black market for substances like chlorofluorocarbons (CFCs), as sales were not restricted. This cost gap between illegal chemicals and their more expensive alternatives provided lucrative profit opportunities [4]. Consequently, these unlawful activities are worsening ozone layer depletion, contributing to climate change, and accelerating global warming, emphasising the urgent need for enhanced international cooperation and robust enforcement mechanisms to combat the illicit trade of these substances.
The PERIVALLON project contribution
The PERIVALLON project (European Union’s – GA No. 101073952) aims to provide and validate novel tools and solutions for improving the operational investigation capabilities for European authorities and relevant security practitioners regarding the identification of illegal trade of ODS and HFCs. The fight against such organised environmental crime is particularly challenging due to the range of items and substances illegally transported, intentionally dumped, and/or traded. Thus, diverse technologies are required for detecting and preventing such activities. These challenges are further compounded by the need to analyse and correlate information and evidence obtained from heterogeneous sources and to consider the different authorities involved in tackling such crimes.
Leveraging Computer Vision to fight the illicit trade of ODS and HFCs
In recent years, computer vision has witnessed remarkable advancements, enabling machines to perceive and understand visual information across a wide range of applications. One pivotal area of focus is object detection, a well-established subject of study within the computer vision community. Lately, the demand for domain-specific datasets is rising, driven by the increasing need for automation, continuous monitoring, and optimisation of critical tasks across several domains.
Object detection for gas cylinders
An interesting and intriguing domain-specific application is related to detecting and analysing cylindrical objects, particularly gas cylinders. This application holds significant potential use cases in various crucial fields, such as monitoring industrial facilities that handle gas cylinders, preventing fires associated with these items, improving warehouse safety and storage efficiency, as well as monitoring potential environmental crimes. Existing datasets in computer vision primarily focus on general object detection [5], [6] and lack the specificity and diversity required for robust gas cylinder detection and identification of their characteristics.
As part of the PERIVALLON ecosystem, advanced AI-based data analysis methods and tools will be developed to understand better the obtained visual content and extract valuable insights through detecting gas cylinder-related objects and recognising their characteristics (material, size, orientation) in online channels (i.e., advertisements in marketplaces).
Due to the lack of a domain-specific dataset, PERIVALLON researchers have created a comprehensive dataset for gas cylinder detection and identification of their attributes in real-world scenarios. The dataset includes images captured in various environments, from industrial facilities to commercial spaces, replicating challenging conditions in outdoor and domestic settings.
Benchmark evaluation
An extensive benchmark evaluation of state-of-the-art object detection models was conducted to verify the dataset’s application value. Three deep learning models were chosen for this purpose, covering a wide range of detector categories, including single-stage, two-stage, and transformer-based architectures: Faster R-CNN [7], YOLOV8 [8], and RT-DETR [9]. Transfer learning was adopted to enhance the models’ performance by pre-training on a generic object dataset, aiding in learning generic features and patterns from a vast array of object classes. As identified during the experiments, the knowledge transfer helps enhance the models’ ability to capture relevant visual cues for more effective localisation of gas cylinders and recognition of their attributes.
Figure 1 Examples of gas cylinders automatic identification in real-world scenarios
Figure 2 Examples of automatic recognition of material, size, and orientation attributes of gas cylinders
[1] Yamaguchi, Shunta. “The nexus between illegal trade and environmental crime.” (2023).
[2] Coulomb, Didier, Jean-Luc Dupont, and Valentin Morlet. “The impact of the refrigeration sector on climate change-35. Informatory note on refrigeration technologies.” (2017).
[3] Unfccc, U. N. Kyoto protocol reference manual on accounting of emissions and assigned amount. eSocialSciences, 2009.
[4] Elliott, Lorraine. “Smuggling networks and the black market in ozone depleting substances.” Hazardous Waste and Pollution: Detecting and Preventing Green Crimes (2016): 45-60.
[5] Lin, Tsung-Yi, et al. “Microsoft coco: Common objects in context.” Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer International Publishing, 2014.
[6] Everingham, Mark, and John Winn. “The PASCAL visual object classes challenge 2012 (VOC2012) development kit.” Pattern Anal. Stat. Model. Comput. Learn., Tech. Rep 2007.1-45 (2012): 5.
[7] Ren, Shaoqing, et al. “Faster r-cnn: Towards real-time object detection with region proposal networks.” Advances in neural information processing systems 28 (2015).
[8] Jocher, Glenn, Ayush Chaurasia, and Jing Qiu. “YOLO by Ultralytics.” URL: https://github. com/ultralytics/ultralytics (2023).
[9] Lv, Wenyu, et al. “Detrs beat yolos on real-time object detection.” arXiv preprint arXiv:2304.08069 (2023).