【摘要】 为加强施工场景下的工程车辆安全监管,针对施工场景下工程车辆易互相遮挡、局部特征与全局特征相似以及场景环境复杂的问题,提出一种基于自适应局部斥力与归一化面积损失的工程车辆目标检测算法。其中自适应局部斥力损失可使待检测工程车辆与其他工程车辆目标框相排斥;归一化面积损失使网络学习集中在面积具有相对较大预测误差的工程车辆上;并结合聚类算法设定更适合工程车辆的锚框。研究结果表明:算法可实现在困难场景下对压路机、挖掘机、装载机3类工程车辆的快速准确检测与识别,具有较高的工程应用价值。更多还原

【Abstract】 In order to strengthen the safety supervision of engineering vehicles under the construction scene, aiming at the problems that the engineering vehicles are easy to obscure each other, the local features are similar to the global features, and the scene environment is complex under the construction scene, an object detection algorithm of engineering vehicles base on the self-adaptive local exclusion loss and normalized area loss was proposed.The self-adaptive local exclusion loss could make the target box of engineering vehicle to be detected to repel with those of other engineering vehicles.The normalized area loss could make the network learning focus on the engineering vehicles with relatively large prediction error of area.Combined with the clustering algorithm, the anchor box being more suitable for the engineering vehicles was set.The results showed that the algorithm could realize the rapid and accurate detection and identification of three types of engineering vehicles, namely road roller, excavator and loader in difficult scenarios, and it has high engineering application value.更多还原