Deer Detection in Thermal Images for Traffic Safety Using Contour Based Histogram of Oriented Gradient Method

Car accidents due to Deer-Vehicle Crashes (DVC’s) are one of the major safety concerns for the driving on rural roads in Europe and North America. Many attempts have been made to reduce the occurrences of these accidents, but none of them has yet proved effective. With the development of infrared imaging technology, using thermal images to identify the presence of deer in night has become possible. A pattern recognition method, Histogram of Oriented Gradient (HOG), can be used to deer detection to reduce such accidents. However, the length of time required to process one image makes the method incapable of real-time deer identification and tracking. Based on the HOG method and a known classifier Support Vector Machine (SVM), this research developed a contour based HOG + SVM method, called CNTHOG method to process thermal images and to detect the presence of deer. Experimental results have demonstrated that this algorithm has the advantages of both high accuracy (up to 94.2%) and short computation time (0.1s) when compared to traditional HOG + SVM method. In this research, further analysis has been performed to evaluate the influence of body postures and occlusions on detection accuracy. By achieving such computation speed and accuracy, this CNT-HOG algorithm proves that deer can be tracked in real-time. By using such a system, drivers on road can be warned of the presence of deer in real time and the frequency of DVC’s can be effectively reduced. KeywordsDeer-vehicle Crashes; Thermal imaging; HOG; Contour