无人驾驶危险交通场景中小目标检测算法研究

Research on Detection Algorithm of Small Targets in Driverless Dangerous Traffic Scenes

无人驾驶车辆对目标检测算法的实时性和检测精度要求很高,尤其是危险交通场景中的小目标检测尚存在完善。针对这个问题,选用YOLOv4作为基础网络,基于组合剪枝策略对YOLOv4进行修剪。为了在剪枝的过程中不降低检测精度,实验通过大尺度训练与网络添加空间金字塔池化来增强深层特征的提取。剪枝策略通过向信道比例因子添加L1正则化来加强信道级的稀疏性促进结构化的剪枝。然后在通道剪枝的基础上融合层剪枝,把比例因子的最小值对应的层裁剪掉,即YOLOv4-Pocket算法。其平均精度(mAP)提高了6.05%,模型空间缩小了99.15%,每帧图像的推理时间缩短了82.45%。实验结果表明,YOLOv4-Pocket模型更适合于无人驾驶汽车的应用场景。

Driverless vehicles require high real-time performance and precision of target detection algorithm,especially for small targets detection in dangerous traffic scene.It has not been completely solved.To solve this problem,we choose YOLOLv4 as the base network and prune it based on the combination prune strategy.In order not to reduce the detection accuracy in the pruning process,the experiment enhanced the extraction of deep features by large-scale training and adding spatial pyramid pooling(SPP)to the network.The pruning strategy enhances channel-level sparsity by adding L-1 regularization to the channel scaling factor to promote structural pruning.Then,layer pruning is integrated on the basis of channel pruning to cut out the layer corresponding to the minimum value of scale factor,namely,YOLOv4-Pocket algorithm.The mean average precision(mAP)is improved by 6.05%,the model space is reduced by 99.15%,and the reasoning time per frame is shortened by 82.45%.The experimental results show that YOLOv4-Pocket model is more suitable for the application of driverless cars.