WildSight | Algotechniq

Motivation

Collisions between vehicles and animals present a significant issue in numerous countries, making it crucial to predict accidents and associated costs at regional and national levels for effective accident prevention (Gren & Jägerbrand, 2019).

Furthermore, a recent bear encounter at a clinic in Niigata Prefecture reflects a worrying trend of bears moving closer to human settlements in Japan. Due to a shortage of acorns, their primary food, and the spread of untamed vegetation, bears are increasingly foraging near human areas. A new term, “shin-sedai kuma” or “new-generation bears,” describes those born near human communities who are less afraid of people. This shift in bear behavior highlights the need for greater caution and a reassessment of human-bear interactions.

Detecting bears and cars by Algotechniq's WildSight.

Objective

The goal of the WildSight project is to improve the accuracy and robustness of wild animal detection in complex field scenarios by using the YOLOv8 algorithm. The focus is on localizing and recognizing wild animals in different challenging environments, including scenes with multiple targets, small targets, or obstructed targets.

Prior Work

Previously, wild animal detection relied heavily on manual observation and feature-based algorithms, such as the HOG (Histogram of Oriented Gradients), and DPM (Deformable Parts Model) algorithms. These methods involved manually designed features and detectors, which were limited by their inability to achieve high accuracy and robustness in complex scenes. More recently, deep learning-based methods, particularly transfer learning techniques, have been used but lacked the real-time performance needed for field applications (Russo et al., 2024).

Benefits

Municipalities and safety authorities can deploy this technology to monitor and manage the presence of wild animals in urban or suburban areas. Early detection of animals near roads, residential areas, or public spaces can help prevent accidents, such as vehicle collisions with large animals, thereby improving public safety.
Other beneficiaries include biologists and researchers involved in wildlife monitoring and conservation. The approach can also be valuable to photographers, drone operators, and others involved in outdoor exploration and wildlife studies who require accurate and efficient animal detection in various environmental conditions.

Enabling Technologies

The key enabling technologies of the new approach include:

YOLOv8 is a new real-time object detection algorithm that builds upon earlier YOLO versions, offering enhanced accuracy and speed, for images and videos.
Data Augmentation Techniques to enhance the training dataset and improve model robustness.

Performance Criteria

The success of the new approach can be quantified through metrics such as precision, recall, mean Average Precision (mAP), and loss function trends during training. Additionally, the model’s ability to perform well in complex scenarios, such as detecting multiple, occluded, or small targets, serves as further evidence of its success.

Neural networks and deep learning are in rapid development; new technologies and increasingly high-performance computing centres allow operations to be carried out that, until a few years ago, were impossible due to excessive processing times. The possibility of exploiting an Edge/Cloud configuration allows speed in training neural networks, ability to operate in real-time, elasticity, and resistance of the entire system in the event of a failure. This configuration helps reduce costs by using less expensive tools in the sensing and edge regions. In this work, an edge computing system was implemented to recognise wild animals such as goats and wild boars. The image recognition problem was therefore addressed by exploiting transfer learning techniques on two state-of-the-art methods: YOLOv5 and EfficientNet. The comparison of the results highlighted the pros and cons of the two methods.

Animal-vehicle collisions (AVCs) involving ungulate species pose a serious problem in many countries, and the prediction of accidents and costs on a regional and national spatial scale is important for efficient accident mitigation. Based on the assumption that AVCs are determined by traffic volume and ungulate population dynamics, this study developed a relatively simple method for calculating and predicting the costs of current and future traffic accidents involving moose, roe deer and wild boar in Sweden. A logistic population model was assumed for all three ungulate species and econometric methods were applied to obtain population growth models based on panel data on traffic accidents, traffic load, hunting bags, hunting licenses and landscape characteristics for each Swedish county and year from 2003 to 2015. The population growth models were used to predict vehicle accidents and costs. The predicted annual discounted costs of AVCs over a 15-year period based on projected ungulate populations and traffic volume fell by 25% from 406 million USD in 2015, however the allocation of costs between ungulates differed. AVCs involving roe deer accounted for the largest share of the costs (54%), but collisions involving wild boar showed the most rapid increase over the study period because of a relatively high estimated growth rate and recent expansion of wild boar populations to several new counties. However, the predicted costs were sensitive to assumptions regarding population dynamics as well as assumptions about future hunting pressure and traffic volume.

Motivation

Objective

Prior Work

Benefits

Enabling Technologies

Performance Criteria

References

2024

2019