基于密集连接卷积神经网络的道路车辆检测与识别算法

Road vehicle detection and recognition algorithm based on densely connected convolutional neural network

针对现有道路车辆检测识别算法中存在的检测精度不高、实时性差以及小目标车辆漏检等问题,提出一种基于密集连接卷积神经网络的道路车辆检测与识别算法。首先,基于YOLOv4网络框架,通过采用密集连接的深度残差网络结构,加强特征提取阶段的特征复用,实现对浅层复杂度较低的特征的利用;然后,在多尺度特征融合网络引入跳跃连接结构,强化网络的特征信息融合和表征能力,以降低车辆漏检率;最后,采用维度聚类算法重新计算先验框尺寸,并按照合理的策略分配给不同检测尺度。实验结果表明,该算法在KITTI数据集上获得了98.21%的检测精度和48.05 frame/s的检测速度,对于BDD100K数据集中复杂恶劣环境中的车辆也有较好的检测效果,在满足实时检测要求的同时有效提升检测精度。

Regarding to the problems of low detection accuracy,poor real-time performance,and missed detection of small target vehicles in existing road vehicle detection and recognition algorithms,a road vehicle detection and recognition algorithm based on densely connected convolutional neural networks was proposed.Firstly,Based on YOLOv4(You Only Look Once version 4)network framework,by adopting the densely connected deep residual network structure,the feature reuse in the feature extraction stage was strengthened to realize the use of features with lower complexity on shallow layers.Then,a jump connection structure was integrated to the multi-scale feature fusion network to strengthen the feature information fusion and expression capability of the network,which reduced the missed detection rate of vehicles.Finally,the dimensional clustering algorithm was used to recalculate the anchor sizes,which were allocated to different detection scales according to a reasonable strategy.Experimental results show that the proposed algorithm achieves the detection accuracy of 98.21% and the detection speed of 48.05 frame/s on KITTI dataset,and it also has a good detection effect for vehicles in the complex and harsh environment of Berkeley DeepDrive(BDD100K)dataset,ensuring required real-time performance and effective accuracy improvement.