Abstract:

In order to solve the problem of insufficient detection performance for vehicles with large scale transformation and overlapping, this paper proposes a single-stage deep convolutional neural network for real-time vehicle detection. For the scale-sensitive problems of vehicle detection algorithms, the deep residual network is used as the feature extraction layer to obtain fine-grained features, and the feature pyramid network is constructed for multi-scale detection; the soft non-maximum suppression is used to linearly attenuate the confidence score to solve the problem of vehicle overlapping so that the vehicle’s missed detection rate can be reduced. At the same time, it pruns the model at channel level to reduce the size of model parameters, saves computing resources and improves the speed of model detection. Experiments on VOC dataset show that the proposed method has high performance in both detection accuracy and detection speed. In terms of detection accuracy, it reaches 87.6%, which is 3.7 percentage points higher than that of YOLOv3, and 9.8 percentage points higher than that of SSD512. In terms of detection speed, the number of detection frames per second reaches 42 f/s, realizing real-time detection of vehicles. In particular, applying the model to the Apollo dataset with complex environment, it has better robustness than YOLOv3.

Key words: vehicle detection, convolutional neural network, residual network, feature pyramid network, network pruning

摘要：

为了解决对于尺度变换较大车辆及遮挡车辆检测性能不足的问题，提出了一种实时车辆检测模型。针对车辆检测算法对于尺度敏感的问题，通过使用深度残差网络作为特征提取层，构建特征金字塔网络用于多尺度检测；利用软化非极大抑制线性衰减置信得分解决车辆遮挡问题，从而降低车辆的漏检率；同时对模型进行通道级裁剪缩减模型参数规模，节省计算资源，提高模型检测速度。在VOC数据集上进行实验，结果表明，提出的方法在检测精度和检测速度上均获得较高的性能。在检测精度上，达到87.6%的准确率,相较于YOLOv3提升了3.7个百分点，相较于SSD提升了9.8个百分点；在检测速度上，每秒检测帧数达到42 f/s，实现了车辆的实时检测。特别地，将模型应用于环境复杂的Apollo数据集，相较于YOLOv3具有更好的鲁棒性。

关键词: 车辆检测, 卷积神经网络, 残差学习, 特征金字塔, 网络裁剪