基于卷积神经网络的多目标实时检测

Multi-target real-time detection based on convolutional neural network

为解决智能驾驶应用场景下以行人与车辆为主,附带骑行电动车的多目标检测存在检测速度不满足实时性的问题,提出一种多目标实时检测方法。通过对YOLO (you look only once)v2卷积神经网络模型进行维度聚类分析以及结构调整等优化举措,行人与车辆检测AP (平均准确率)值分别为71%和81%,检测速度为50帧/s。实验结果表明,该方法与目前先进目标检测方法相比,在准确率相差5%以内的前提下大幅提高检测速度,实现了实时性检测的目标。

To solve the problem that the detection speed of multi-objective detection of pedestrians and vehicles with attached electric vehicle riding in real-time is not satisfied with the real-time performance in smart driving applications,a multi-objective realtime detection method was proposed.Dimensional clustering analysis and structural adjustment and other optimization measures of YOLO(you look only once)V2 convolution neural network model were carried out.The average pedestrian and vehicle detection AP(average accuracy)values are 71%and 81%,respectively.The detection speed is 50 frames/s.Experimental results show that compared with the current advanced target detection method,the proposed method greatly improves the detection speed on the premise of a difference of less than 5%in accuracy,realizing the goal of real-time detection.