【摘要】 在实际场景中，车辆目标往往会被其他车辆、建筑物等对象遮挡，背景也可能非常复杂，为了保障检测精度，提出一种基于融合卷积神经网络的车辆多目标检测方法。采用激光雷达采集车辆目标图像，将采集的车辆行驶图像根据其车道线特征划分为两侧区域，将车道线以内的区域作为车辆多目标检测初始感兴趣区域(ROI),在ROI中采用车底阴影假设区域分割法获取车辆检测目标的假设区域。在原始卷积神经网络的基础上作进一步优化，设计可变形卷积神经网络(DF-R-CNN)模型，将得到的假设区域作为网络模型所需的车辆多目标检测候选区域，通过该模型实现车辆多目标的精准检测。实验结果表明，所提方法的召回率最高值达到了85%,损失函数最低值约为1.8,说明其具有较高的检测精度和检测效果。更多还原

【Abstract】 In practical scenarios, vehicle targets are often obstructed by other vehicles, buildings, and other objects, and the background may also be very complex. To ensure detection accuracy, a vehicle multi-target detection method combining convolutional neural networks and LiDAR is proposed. Using LiDAR to capture vehicle target images, the collected vehicle driving images are divided into two sides based on their lane line characteristics. The area within the lane line is used as the initial region of interest(ROI) for vehicle multi-target detection. In the ROI, the vehicle bottom shadow hypothesis region segmentation method is used to obtain the hypothesis region for vehicle detection targets. On the basis of the original convolutional neural network, further optimization is carried out to design a deformable convolutional neural network(DF-R-CNN) model. The assumed region obtained is used as the candidate region for vehicle multi-target detection required by the network model, and accurate detection of vehicle multi-targets is achieved through this model. The experimental results show that the highest recall rate of the proposed method reaches 85%, and the lowest loss function value is about 1.8, indicating that it has high detection accuracy and effectiveness.更多还原