Ensemble deep learning approach for traffic video analytics in edge computing

Video analytics is the new era of computer vision in identifying and classifying objects. Traffic surveillance videos can be analysed to using computer vision to comprehend the road traffic. Monitoring the real-time road traffic is essential to control them. Computer vision helps in identifying the vehicles on the road, but the present techniques either perform the video analysis on the cloud platform or the edge platform. The former introduces more delay in processing while controlling is needed in real-time, the latter is not accurate in estimating the current road traffic. YOLO algorithms are the most notable ones for efficient real-time object detection. To make such object detections feasible in lightweight environments, its tinier version called Tiny YOLO is used. Edge computing is the efficient framework to have its computation done on the edge of the physical layer without the need to move data into the cloud to reduce latency. A novel hybrid model of vehicle detection and classification using Tiny YOLO and YOLOR is constructed at the edge layer. This hybrid model processes the video frames at a higher rate and produces the traffic estimate. The numerical traffic volume is sent to Ensemble Learning in Traffic Video Analytics (ELITVA) which uses F-RNN to make decisions in reducing the traffic flow seamlessly. The experimental results performed on drone dataset captured at road signals show an increase in precision by 13.8%, accuracy by 4.8%, recall by 17.4%, F1 score by 19.9%, and frame rate processing by 12.8% compared to other existing traffic surveillance systems and efficient controlling of road traffic.