AbstractsEngineering

Depth estimation plays an important role in three-dimensional computer vision area. Its recent development focuses on real-time application. To be able to provide depth map for real-time applications like pedestrian detection and intelligent vehicles, three challenges must be overcome: 1. Real-time processing speed; 2. Insensitive to brightness change; 3. Clear boundaries and smooth surface. The thesis first describes the major steps of depth estimation, then many commonly used methods are reviewed. From them and other related real-time method, we found that the iteration based semi-global matching with mutual information has much potential to be improved. Based on that, the thesis proposed a method with two layers to provide depth map for pedestrian detection. The low resolution layer’s task is to produce a coarse depth map as quick as possible, then it helps to produce accurate mutual information distribution without iterations. The Full-resolution layer applies semi-global matching with two optional simplification schemes to speed up. The proposed method is implemented on both CPU and GPU. Experimental results and evaluation shows that it has great insensitivity to brightness changes and achieves real-time processing speed while maintaining a comparable performance with state-of-the-art real-time depth estimation methods.