, Professor, UMD-Mechanical & Industrial Eng
Deer-vehicle collisions (DVC) are one of the most serious traffic issues in the United States. To reduce DVCs, this research developed a system using a contour-based histogram of oriented gradients algorithm (CNT-HOG) to identify deer through processing images taken by thermographic cameras. The system is capable of detecting deer in low-visibility conditions. Two motors are applied to enlarge the detection range and make the system capable of tracking deer by providing two degrees of freedom. The main assumption in the CNT-HOG algorithm is that the deer are brighter than their background in a thermal image. The brighter areas are defined as the regions of interest, or ROIs. ROIs were identified based on the contours of brighter areas. HOG features were then collected and certain detection frameworks were applied to the image portions in the ROIs instead of the entire image. In the detection framework, a Linear Support Vector Machine classifier was applied to achieve identification. The system was tested in various scenarios, such as at a zoo and on roadways in different weather conditions, and the influence of the visible percentage of a deer body and the posture of a deer on detection accuracy was measured. Results have shown that this system can achieve high-detection accuracy (up to 100 percent) with fast computation speed (10 Hz). Achieving such a goal will help to decrease the occurrence of DVCs on roadsides.