时间: 2025-05-08 | 次数: |
王妍玮,张佳宇,陈凯云,任春平. 将异步事件流转换为网格表示的方法研究[J].河南理工大学学报(自然科学版), doi: 10.16186/j.cnki.1673-9787.2024070033.
WANG Y W,ZHANG J Y,CHENG K Y,REN C P. Methods Research for Converting Asynchronous Event Streams into Grid Representations [J]. Journal of Henan Polytechnic University( Natural Science), doi:10.16186/j.cnki.1673- 9787. 2024070033.
将异步事件流转换为网格表示的方法研究(网络首发)
王妍玮,张佳宇,陈凯云,任春平
(黑龙江科技大学 机械工程学院,黑龙江 哈尔滨 150022)
摘要:目的 针对异步事件流的复杂性和稀疏性带来的数据分析复杂化以及存储计算效率低等问题。方法 提出一种将异步事件流转换为网格表示的方法,首先采用Dirac脉冲以函数的形式替代每一个事件,并将其表示为一个事件场集合,根据事件的张量特性为每个丢失最多一类信息的事件分配一个该类别的平均测量值,减小了计算量的同时保留了事件的高动态分辨率;然后选择能直接利用数据来寻找最佳的测量函数候选项的多层感知器MLP,代替原有手动选择的聚合核函数,在ECTResNet中进行卷积,通过定期采样以降低维度并保留关键信息进行量化处理,再次经过卷积后将事件网格化,通过连续的三维空间离散产生一个固定大小的网格,最后将网格化的事件流转换为可深度学习的网格表示形式作为网络的输出。结果 实验在两个公开数据集N-Cars和N-Caltech101上进行分析,经网络转换后的输出表示识别准确率可达97.07%和87.72%,比事件尖峰张量的方法提高了10.09%和11.44%。结论 实验表明,将异步事件流转换为网格表示可以更好地适应深度学习模型,提高事件处理和识别的准确性和效率,并支持端到端的表示学习,可在传感器数据处理和事件识别领域中得到广泛的应用。
关键词: 事件相机;异步事件流;深度学习;卷积神经网络
中图分类号:TP242
doi: 10.16186/j.cnki.1673-9787. 2024070033
基金项目: 国家自然科学基金资助项目(52204131);黑龙江省重点研发项目(GA23A910);黑龙江省科研基本业务费(2022-KYYWF-0527)
收稿日期:2024-07-05
修回日期:2024-07-25
网络首发日期:2025-05-08
Methods Research for Converting Asynchronous Event Streams into Grid Representations (Online)
WANG Yanwei, ZHANG Jiayu,CHEN Kaiyun,REN Chunping
(College of Mechanical Engineering, Heilongjiang University of Science &Technology,Harbin 150022,Heilongjiang, China)
Abstract: Objective To addressing the complexity and sparsity of data analysis, as well as the low storage and computational efficiency,which caused by asynchronous event streams. Methods A method is proposed to convert asynchronous event streams into grid representations.Firstly, Dirac pulse is used to replace each event by a function form, which is represented as an event field set. In order to decreasing the computational cost and holding the high dynamic event resolution, Assigning to each event that loses the most information of a category an average measurement that category based on the tensor properties of the event. Secordly, the multi-layer perceptron MLP, which can directly use the data to find the best candidate value, is selected by measurement function instead of manually selected aggregate kernel function. It convoluted in ECTResNet to reducing dimensionality and keep key information for quantization through regular sampling. Finally, according to the reconvolution , events are gridded, and a fixed-size grid is generated through continuous three-dimensional space discretization, the final gridded event stream is converted into a grid representation ,which can be used for deep learning as the network output. Results The results are acquired by two public datasets N-Cars and N-Caltech101, and the output representation recognition accuracy of the transformed datasets reach 97.07% and 87.72%, which are higher than the event spike tensor method, which are 10.09% and 11.44% . Conclusion Experiments show that converting asynchronous event streams into grid representations can better adapt to deep learning models, it improve the accuracy and efficiency, which events are processed and recognized. And it support end-to-end representation learning, which can be widely used in the field of sensor data processing and event recognition.
Key words:event camera; asynchronous event streams; deep learning; convolutional neural networks
CLC: TP242