Adaptive density based data mining technique for detection of abnormalities in traffic video surveillance

Abstract

Detection of abnormal events in a traffic scene is a highly challenging task due to vast field of view, continuous stream of video data, various object interactions and complex events in Video Surveillance. Hence, this research proposes novel schemes using machine learning approach to detect abnormal events such as illegal U-turn, presence of pedestrian in driving region, wrong side driving and frequent lane change. Recently, Density Based Spatial Clustering of Applications with Noise (DBSCAN) is a popular method that has been used for clustering the trajectory datasets. The existing Density Based Clustering approach used for Abnormal detection in traffic scene uses random selection of cluster radius (Eps) and minimum points (minpts) needed to form a cluster. This random selection is time consuming and inefficient clustering results in accuracy reduction in abnormal detection. So, Adaptive Density based Spatial Clustering of Applications with Noise (ADBSCAN) is proposed for the detection of abnormal events based on spatial temporal information relating to individual objects which determines the optimal values for the cluster radius (Eps) using the slope calculation of the K-d plot. Gaussion Mixture Model (GMM) is used for obtaining the moving foreground regions and region-based tracking is used for the identification of the objects in successive frames. The centroid of the region is calculated using image moments. If there is an occlusion between the vehicles then vehicle identification number (Id no) is used to differentiate them. The main advantage in this technique is clustering/labelling the normal pattern without the help of manual intervention. The effectiveness of ADBSCAN is experimentally evaluated using a real time benchmark video traffic dataset and it found that it gives better accuracy in detecting anomalies than the state-of-the-art techniques.

A New Density-Based Clustering Method Considering Spatial Distribution of Lidar Point Cloud for Object Detection of Autonomous Driving

“…The parameters of DAC are shown in Table 1, following the selection principle of parameters from Section 4.2. When designing the SDSC method, the detected region was divided into three sub-regions based on the density histogram, and each of the parameters is calculated by k-dist [42] as shown in Table 2. Time, weather and other conditions affect the movement of people, vehicles etc.…”