Aresearch on wide area traffic event detection systems by information fusion through ubiquitous sensor network

在本项目中，我们的目标是通过采用不同类型的传感器，如视觉传感器和超声波传感器的网络开发的广域交通事件检测系统的驾驶员的援助。我们安装了事故检测系统来收集事故的图像数据，并对所获得的事故图像进行调查，以揭示引发事故的机制。通过研究发现了一个有趣的机制，即下游车辆的小断裂行为所引起的冲击波在传播到上游车辆的过程中被放大，冲击波应该是高速公路事故的主要因素，因此我们开发了传感器融合网络系统，以尽快检测这种冲击波的传播。传感器融合网络系统采用两公里长的图像传感器和超声波传感器，从传感器收集的数据将被汇总用于驾驶员辅助。图像传感器能够 关于我们 获取每辆车的轨迹和每车道的车辆数，而超声波传感器能够获取每辆车的速度和每车道的车辆数。车辆的速度可以从图像传感器的车辆的轨迹计算。通过总结传感器的速度和车辆计数，我们开发了一种算法来预测由冲击波传播引起的每个位置的状态转换。该算法通过划分时空立方体来定义盒子，时空立方体代表每个时间每个位置的速度和车辆数量，并通过学习6个月的状态转换数据来建模状态转换。结果表明，该算法能够以85-90%的成功率预测每个时刻每个地点的状态。最后，我们通过雇用40名由不同年龄段的男性和女性组成的监控驾驶员来检验驾驶员支持系统的效果。通常，这种基于模式识别的系统具有不确定性，例如未检测或错误检测。通过降低检测的阈值水平，漏检测增加而误检测减少。通过提高检测的阈值水平，漏检减少而误检增加。因此，优化模式识别参数以使系统最适合驾驶员是重要的。通过驾驶模拟，确定了冲击波检测系统的最佳参数，并通过六个月的交通数据对优化后的系统进行了验证。少

In this project, we aimed at development of wide area traffic event detection systems for driver's assistance by employing network of different kinds of sensors such as vision sensors and supersonic wave sensors. We installed incident detection system to collect image data of accidents, and investigated such acquired accident images to reveal mechanisms arousing accidents. By the investigation, an interesting mechanism was revealed as such that shock waves caused by a small breaking behaviors at downstream traffic are amplified during propagating to upstream traffic, and the shock waves should be a major factor for the accidents on highways.We thus developed sensor fusion network systems to detect propagation of such the shockwaves as soon as possible. The sensor fusion network systems employ image sensors and supersonic wave sensors for two kilometers long, and the collected data from the sensors will be summarized for the purpose of driver's assistance. The images sensors are able to … More acquire trajectory of each vehicle and vehicle count at each lane, while the supersonic wave sensors are able to acquire velocity of each vehicle and vehicle count at each lane. Velocities of vehicles can be calculated from the trajectories of vehicles of image sensors. By summarizing those velocities and vehicle counts from the sensors, we developed an algorithm to predict the state transition at each location caused by shock wave propagation. The algorithm defines boxels by dividing time-space cubes that represents velocity and vehicle count at each location at each time, and it modeled the state transition by learning six months data of the state transition. As a result, our algorithm was able to predict the state of each location at each time in 85-90% successful rate.Finally, we examined the effect of driver's support system by employing forty monitor drivers consisting of men and women of wide ages. In general such the pattern recognition based system has uncertainty such as miss detections or false detections. By lowering threshold level for the detection, miss detections increase while false detection decrease. By elevating threshold level for the detection, miss detections decrease while false detection increase. Therefore, it is important to optimize parameters for pattern recognition in order to make the system most agreeable for the drivers. By the driving simulation, we determined the most appropriate parameters for the shock wave detection systems, and we validated the optimized system by applying to six months traffic data. Less