基于车型聚类的交通流参数视频检测

单目摄像机成像丢失深度信息,且PTZ (Pan/Tilt/Zoom)摄像视频场景多变,导致交通流参数提取误差较大. 提出了一种基于车型聚类的交通流参数检测方法. 在改进的摄像机自标定成像模型中,提取PTZ 参数变化下的透视投影不变量"伪形状特征', 对其进行基于贡献率算法的车型聚类分析,以车型均高代替实际高度,获取车辆的长宽, 进而计算道路空间占有率,并提升车速检测精度. 测试表明实时性较高,车型聚类自适应于不同场景,平均准确度为96.9%,车长计算精度优于90%.

Detecting Traffic Parameters Based on Vehicle Clustering from Video

Traditional methods of traffic parameter extraction often result in large errors because of the lost of depth information in monocular camera imaging and the frequent changes of the PTZ (Pan/Tilt/Zoom) camera video scenes. In this paper, a video-based traffic parameter extraction method using vehicle clustering is proposed. Built on the new method is a modified camera self-calibration imaging model in which the pseudo-form feature of vehicles projection is described. For purposes of obtaining vehicle length and width, the imaging model is improved by introducing a contribution rates algorithm for vehicle clustering, and using the average vehicle height instead of the actual height in the model. Thus space occupancy can be obtained, and speed detection accuracy can be increased. Test results show that: high real-time performance is achieved; vehicle clustering algorithm is adaptive in various scenes, and has an average accuracy of 96.9%; the accuracy of vehicle length estimation is greater than 90%.