复杂交通环境下车辆视觉检测关键问题的研究

Study on video-based vehicle detection is the preconditions and foundation of self-guided vehicles and driver-assistances systems. How to improve processing rate and detection accuracy of vehicle detection system under outdoor complex environment is the key to practical application of vehicle detection technology. First, the project adopt an image analysis algorithm to determine vehicle candidate regions, and select the most suitable model from multi-view detection modes based on geometric model of each candidate region; Second, both image global structure and local detail features are extracted to describe each candidate region; Finally, a fast feature pyramid is established, the image features are computed as octave-spaced scale intervals are sufficient to approximate features on a finely-sampled pyramid, which can determine whether there is a vehicle in each vehicle candidate region quickly, and achieve high accuracy of vehicle detection. The performance of the propose system will be validated by theoretical analysis and experiments. The ultimate goal of the project is to lead to better performance and faster processing rate comparing with the current state-of-the-art methods, and increase the detection rate more than 5%-10%. Based on the project, a novel object detection framework will be established, the processing rate and detection accuracy of the vehicle detection system will be improved, and the project will promote the practicability of the vehicle detection technology.

基于视频的车辆检测研究是无人驾驶和辅助驾驶系统的前提和基础。如何提高车辆检测系统在室外复杂环境下的检测能力、提高系统的处理速度是车辆检测技术实用化的关键。本项目拟通过图像分析算法初步确定车辆的候选区域，并根据车辆候选区域的几何模型从多视角检测模型中挑选最适合的检测模型；其次，融合图像的全局结构与局部细节特征对车辆候选区域进行描述；最后，建立快速图像特征金字塔模型，利用从标准层级中提取的图像特征，通过近似算法扩展到更多的层级上去，从而快速地判断多尺度图像空间的车辆候选区域中是否包含有车辆，实现高精度的车辆检测。本项目拟采用仿真实验和理论分析相结合的方法评价系统的性能。最终目标是，提出的算法实现比传统车辆检测方法更精确、更快速的检测效果，达到将检测率提高5%-10%以上的预期目标。通过本项目的研究，将提出新型的车辆检测方法，提高车辆检测速度与精度，为推进车辆检测系统的实用化奠定基础。

基于视频的车辆检测研究是无人驾驶和驾驶辅助系统的前提和基础。如何提高车辆检测系统在室外复杂环境下的检测能力、提高系统的处理速度是车辆检测技术实用化的关键。本申请项目通过图像分析算法初步确定车辆的候补区域，并通过车辆候补区域的几何模型从多视角检测模型中挑选最适合的检测模型;本项目建立了快速金字塔检测模型，利用从图像金字塔的某个层级中提取到的图像特征，通过近似算法扩展到更多的层级上去，从而快速地判断多尺度空间的车辆候补区域中是否包含有车辆，完成高精度的车辆检测。本项目取得的主要成果有：1）分别在KITTI等国际公开车辆检测库中进行实验仿真，实验结果证明了本算法与现有的车辆检测算法相比检测精度提高了5%以上；2）建立了中国交通图像数据库，人工标注不同天气环境的交通图像达15万帧；3）建立了功能完备的交通图像分析嵌入式平台；4）在本项目的资助下，共发表学术论文2篇，项目申请人以第一作者在IEEE Trans 期刊上发表论文1篇；5）在本项目的资助下，培养硕士研究生4名。