Research On Event Based Moving Object Detection Algorithm

Moving object detection has always been a hot topic in the field of computer research and application,widely used in fields such as video surveillance and autonomous driving.With the iteration and upgrading of computer technology,the accuracy and speed requirements of object detection are constantly improving,especially in complex scenes such as high-speed moving scenes,scenes with strong or dark lighting intensity,etc.At the same time,the resource consumption and computational time of the algorithm are constantly increasing,and the effectiveness of moving object detection still needs to be further improved.In response to this issue,this paper introduces the emerging visual sensor event camera to solve the problem of moving object detection in high-speed or dim scenes,and proposes two moving object detection algorithms based on event data.The specific research content is as follows:(1)We have created pixel level composite datasets and VOT2013 car events composite datasets.By analyzing the principle of event data generation,an event integration model and an event cumulative update model were established.The reconstruction of event frames was achieved from the perspectives of segmentation and asynchronous output,respectively,and visual images were provided.(2)An event based improved spectral clustering algorithm for moving target detection is proposed.When improving the spectral clustering algorithm,consider the combination of values and directions,combine the cosine distance and Manhattan distance to generate the fusion distance,and establish a more accurate similarity matrix.Using the clustering results of some event data as guidance,and based on reconstructing the relationship between event frames,calculating the clustering results of other data can reduce the required computational time consumption and improve the efficiency of moving object detection.Adaptively obtain the number of clusters to avoid subjective errors caused by manual settings.The detection accuracy of this algorithm on multiple datasets is over 80%,which is greatly improved compared to other commonly used clustering algorithms.The experimental results show that the event based improved spectral clustering algorithm is feasible for moving target detection.(3)A convolutional neural network based on event data was designed.By connecting the reconstructed event frames to the input layer and updating feature information based on the newly incoming event frames,an event based convolutional neural network was implemented.The event based convolutional neural network object detection algorithm performs well on multiple datasets,with an accuracy of over 90%.Experiments have shown that convolutional neural network detection algorithms can also achieve more accurate detection when facing the problem of overlapping and large targets.(4)A probability joint and color joint detection algorithm based on event and image was proposed.After achieving the detection of event data and standard images separately,the final target detection result is obtained based on probability joint detection to compensate for the limited information of the event and the motion blur of the image.In addition,during color joint detection,the RGB color space is converted into the HSV color space for color comparison,and binarization is performed for specific color targets.Synchronize the target position predicted by the convolutional neural network into the binary frame and calculate the proportion of white pigment.Finally,the detection effect is determined by the relationship between the proportion and a certain threshold.The experimental results show that the probability joint detection algorithm can provide reliable detection results in scenarios where the target has different motion speeds,and the color joint detection algorithm can perform well in specific target extraction tasks.