【摘要】 道路目标的识别与定位一直是无人驾驶和视频监控领域中的重要课题之一。以道路监控影像中的车辆为目标,提出一种端到端卷积神经网络车辆检测模型(HyperLocNet)。该模型首先利用区域生成网络(RPN)产生初始候选框,采用动态概率定位模型,给出目标在以初始候选框为中心的搜索区域内的精确位置,提高定位模型的稳定性和收敛性。动态概率模型通过输出搜索区域包含目标的概率,提供丰富的信息用于精确定位。在此基础上,HyperLocNet联合训练目标定位和识别任务实现端到端检测,提高定位精度和检测效率,并在监控视频中采集的道路目标数据集上进行试验,具有很好的检测性能,并且能够达到13帧/s的检测速度,具有实时处理的潜力。更多还原

【Abstract】 Traffic object recognition and location have always been one of the most important topics in the field of unmanned driving and video surveillance. In this paper,an end-to-end convolution neural network vehicle detection model(HyperLocNet) was proposed for vehicle detection in traffic images surveillance. Firstly,the region proposal network(RPN)was used to generate initial candidate boxes,and the dynamic probabilistic location model was used to provide the exact location of the target in the area centered on the initial candidate box,so as to improve the stability and convergence of the location model.The dynamic probabilistic model provides rich information for accurate positioning by outputting the probability of the target in searching area. On this basis, HyperLocNet integrated training object location and recognition task to achieve end-to-end detection,improved positioning accuracy and detection efficiency,carried out experiments on traffic object data collected in surveillance video. The proposed model has good detection performance,achieves the detection speed of 13 frames per second,which has the potential of real-time processing.更多还原