A Hierarchical Edge-Cloud Framework for Collision Aware Video Filtering in Traffic Surveillance Systems

Abstract

The high bandwidth consumption in the network, which is caused by the large-scale vehicular sensing systems, which are characterised by continuous video captures, but safety hazardous situations such as vehicle-vehicle accidents are not common in frequency. In this paper, a hierarchical edgecloud framework has been proposed in collisionaware video filtering as a result of mitigating this inefficiency by transmitting less data and still being able to capture valuable information. The proposed system performs lightweight semantic interaction extraction on the edge and offloads recall oriented collision-based reasoning and aggregation to the cloud enabling selective uploading and resulting uploading of videos rather than raw video streams. The results of the experiments in the held-out test set show that the hierarchical approach has the capacity to achieve a video-level recollection rate of 0.8621 with a reduction of 40.38 percent in the bandwidth usage relatively to a stream-all baseline, which transmits all videos. In contrast, edge-only heuristic has a higher reduction of bandwidth of 86.90 but only retrieves 17.24 percent of all the videos with collisions, a fact that presents the drawbacks of local decision making. Event-level classification is associated with a moderate accuracy, however, once the amount of interaction events per video increases as a whole, performance of the system becomes efficient, yielding the F1-score of 0.6329. Another ablation study also suggests that relative velocity properties also increase recall and bandwidth performance although directional cues do not increase these aspects significantly. These findings indicate that filtering does not require high accuracy of events. Instead, the bandwidth conscious vehicular video analytics can be realized by remembering based semantic compression at the edge with cloud-based aggregation as an implementation of a scalable and practical solution.