Extracting Key Traffic Parameters from UAV Video with On-Board Vehicle Data Validation

Abstract

The advantages of UAV video in flexibility, traceability, easy-operation, and abundant information make it a popular and powerful aerial tool applied in traffic monitoring in recent years. This paper proposed a systematic approach to detect and track vehicles based on the YOLO v3 model and the deep SORT algorithm for further extracting key traffic parameters. A field experiment was implemented to provide data for model training and validation to ensure the accuracy of the proposed approach. In the experiment, 5400 frame images and 1192 speed points were collected from two test vehicles equipped with high-precision GNSS-RTK and onboard OBD after completion of seven experimental groups with a different height (150 m to 500 m) and operating speed (40 km/h to 90 km/h). The results indicate that the proposed approach exhibits strong robustness and reliability, due to the 90.88% accuracy of object detection and 98.9% precision of tracking vehicle. Moreover, the absolute and relative error of extracted speed falls within ±3 km/h and 2%, respectively. The overall accuracy of the extracted parameters reaches up to 98%.

Urban Traffic Monitoring and Analysis Using Unmanned Aerial Vehicles (UAVs): A Systematic Literature Review

“…Nevertheless, these aspects are not discussed in the papers selected for this analysis because they are strictly focused on describing the algorithms for traffic monitoring. As stated in [82], the requirements for real-time traffic management and control generated broad attention in the field of traffic monitoring and new frameworks for low-altitude UAV systems were developed in many countries.…”

“…For vehicle detection, some of the works used conventional computer vision techniques that are focused on feature extraction, such as interest point detection (Shi-Tomasi features) [73], scale invariant feature transform (SIFT) [67,70], histogram of oriented gradients (HOG) features [58,76], local binary patterns (LBP) [49], Viola-Jones object detection scheme [58,76], Haar-like features [56] together with classifiers like support vector machine (SVM) [81], AdaBoost classifier [49,58,76], or k-means clustering [70]. Moreover, for fully automatic techniques of tracking, traditional motion-based methods can be identified, e.g., optical flow (e.g., Kanade-Lucas algorithm) [73][74][75]84,86], background subtraction [61,74,75,77,80], particle filter [28,61,83], correlation filter [68], Kalman filter [65,75,78,82,87,92].…”

“…In addition to these classic methods, deep learning-based methods have been developed using two-step detectors like deep neural network (DNN) [70], RetinaNet [91], convolutional neural network (CNN) [28,79], Faster R-CNN [66], fully connected neural network (fcNN) [70], or one-step detectors: You only look once (YOLO) [28,70,78,82,87] and single shot multibox detector (SSD) [92]. These studies showed that deep learning-based methods are more effective than traditional computer vision techniques in traffic video analysis [92].…”

Legal analysis of implementation of the traffic accident monitoring system

“…At the same time, it is proved that the use of unmanned aerial vehicles (hereinafter -UAVs) as a video surveillance camera for traffic condition provides accuracy of 80%, and the use of stationary video cameras for traffic control has 50-75% accuracy. Besides, UAVs combine the capabilities of both stationary and mobile traffic detectors (Shan et al, 2021). With good visibility from above (without clouds, high-voltage cables and good lighting), UAVs provide the ability to collect more data, with greater accuracy and speed in relation to traditional approaches to recording/registering accidents.…”

FRCNN-Based Reinforcement Learning for Real-Time Vehicle Detection, Tracking and Geolocation from UAS

“…The ability to locate and follow vehicles is significant for security and reconnaissance applications as well as for Intelligent Transportation Systems. Recently, there has been an expanded utilisation of automated airborne vehicles (UAVs) or drones for a reconnaissance due to their ability to observe far-off scenes [5,6]. However, with an increasing number of applications many challenges have appeared.…”