基于时空关联度的视频车祸识别算法

Video Car Accident Recognition Algorithm Based on Spatio-temporal Correlation Degree

摘要:

为了从一大段车祸视频中快速获取车祸发生时的那一小段视频，从中准确识别车祸视频中车辆碰撞的局部信息，从而尽快通知相关部门采取营救措施，避免二次事故的发生，实现快速辅助救援，提出了时空结合的车祸识别方法。通过对车祸视频进行局部特征点的提取和描述，以及对车祸视频进行局部特征时空角点检测，从而实现对车祸视频进行时空特征的提取。经特征提取后可以获得车祸视频的3DSIFT、STIP和3DHOG三种特征，对这三种特征分别进行不同组合的串联融合，融合后可以得到5种串联结果，对其串联特征进行K-MEANS聚类和KNN识别，从而实现车祸视频的识别。实验结果表明，利用时空关联的算法进行识别，能够准确找到车祸发生的片段，可以提高车祸识别精度。对比传统算法所得到的33%识别率，基于时空关联度的视频车祸识别算法可提高车祸识别精度至62%。

Abstract:

In order to quickly obtain a small video of a car accident from a large section of car accident video, the local information of the vehicle collision in the car accident video is accurately identified, so that the relevant department is notified as soon as possible to take rescue measures to avoid the occurrence of a second accident and achieve rapid assistance. Rescue, put forward a combination of space-time car accident identification method. Through the extraction and description of the local feature points of the car accident video, and the detection of the temporal and spatial corners of the local features of the car accident video, the temporal and spatial features of the car accident video are extracted. After the feature extraction, 3DSIFT, STIP and 3DHOG features of the car accident video can be obtained. These three features are respectively combined by different combinations of tandem fusion. After merging, 5 tandem results can be obtained, and K-MEANS clustering of the tandem features is performed. And KNN recognition to realize the recognition of car accident video. The experimental results show that using the algorithm of spatial-temporal correlation to identify the car accident can accurately find the fragments of car accidents, which can improve the accuracy of car accident recognition. Compared with the recognition rate of 33% obtained by the traditional algorithm, the video traffic accident recognition algorithm based on spatio-temporal correlation can improve the recognition accuracy of the crash to 62%.