立体区域卷积神经网络(Stereo R-CNN)算法具有准确、高效的特点,在一定场景下的检测性能较好,但对于远景目标的检测仍有一定的提升空间。为了提升双目视觉算法的车辆检测精度,提出一种改进的Stereo R-CNN算法。该算法将确定性网络(DetNet)作为骨干网络,以增强网络对远景目标的检测;针对左右目视图的潜在关键点,建立了左右视图关键点一致性损失函数,以提高选取潜在关键点的位置精度,进而提高车辆的检测准确性。在KITTI数据集上的实验结果表明,本算法的性能优于Stereo R-CNN,在二维、三维检测任务上的平均精度提升了1%~3%。

Stereo region-convolutional neural networks (Stereo R-CNN) algorithm has the characteristics of accuracy and efficiency. It has better detection performance in certain scenes, but there is still room for improvement in the detection of distant targets. In order to improve the vehicle detection accuracy of the binocular vision algorithm, an improved Stereo R-CNN algorithm is proposed in this paper. The algorithm uses deterministic network (DetNet) as the backbone network to enhance the network's detection of long-term targets; for the potential key points of the left and right eye views, the consistency loss function of the key points of the left and right views is established to improve the location accuracy of the potential key points, and then improve the accuracy of vehicle detection. Experimental results on the KITTI data set show that the performance of the algorithm is better than Stereo R-CNN, and the average accuracies of two-dimensional and three-dimensional detection tasks are improved by 1%-3%.