Abstract:
Animal-vehicle collision (AVC) is a significant global issue that causes considerable
loss of life and property damage. Real-time animal detection using computer vision
techniques and machine learning algorithms is considered to be the most effective way
to prevent AVC. This study presents a review of the state of art methods used for
animal detection and prevention of AVC using image processing techniques. Histogram
of Oriented Gradients (HOG), Local Binary Patterns (LBP), and Haar features are
the commonly used methods for object and animal detection. However, challenges
such as image scale and viewpoint variability, background clutter, lighting conditions,
image quality, and occlusion lower the accuracy and effectiveness of these methods.
Various classifiers such as K-way logistic regression, support vector machine (SVM),
and K-nearest neighbors (KNN) are used to overcome these problems and evaluate
features. LBP-AdaBoost and HOG features showed better results than others, with a
detection rate of 91%, when extracted from regions of interest (ROIs). The performance
of the system is further improved when combined with HOG-SVM. Thus, HOG and
LBP features can be useful for detecting animals despite the defects in images. The
best method for AVC detection is HOG-SVM with a detection accuracy rate of 92%.
DCNN, HOG-AdaBoost, and LBP-AdaBoost showed an accuracy rate of 91%, 84%, and
82% respectively. The worst performance was seen when the Haar-AdaBoost method
was used (79%). However, the effectiveness of each technique depends mainly on the
accuracy of real-time animal detection, the signal transmission speed from the system to
the driver, and the vehicle speed. Comparative analysis can be carried out considering
animal detection for methods used in foggy weather conditions and the effect of speed
variations of the vehicle on the results. Expanding the scope of detection to cover a
broader range of animal species is proposed as a future direction of AVC systems.
