DroneTrafficZA2020: A sample dataset of N4 freeway traffic adjacent to Engineering 4.0 in South Africa

Dataset associated with the primary article: Real-time traffic quantization using a mini edge artificial intelligence platform

A video sample captured by a DJI Mavic Air on the 28th of July at 11:30 of the N4 freeway located in Pretoria, South Africa, adjacent to the Engineering 4.0 campus. The high-definition video (4K, 29.97fps) is 18 minutes and 45 seconds in length (2020718_Mavic_Air_4K.mp4). The second video is a cropped version (1220x560 px) used for the analysis (20200728_Mavir_Air_Cropped.mp4). The analysis involves: (1) generating a ground truth dataset of every vehicle (either a "car" or "truck") entering the frame; only vehicles travelling West (moving right-to-left) toward Hatfield is considered. The dataset contains both the frame number the vehicle is first observed and its associated class, and (2) automatic object detection and counting of vehicles as processed by OpenDataCam. Both the ground truth (groundTruth.csv) and inference (OpenDataCam.csv) dataset is provided in a CSV file format. The line counter configuration file from OpenDataCam (line_counter_config.json) is provided to aid reproducibility. The Python script (ProcessOpenDataCam.py) to process the data is also provided.
OpenDataCam, with the correct threshold settings of the AI network, manages to count vehicles to within a 5% accurarcy.

Summary of files:
+ 2020718_Mavic_Air_4K.mp4 - 4K resolution video file captured by the drone
+ 20200728_Mavir_Air_Cropped.mp4 - Cropped 4K video centered about the freeway (1220x560 px)
+ groundTruth.csv - ground truth data of binary vehicle classification
+ OpenDataCam.csv - vehicle counting and detection output data provided by OpenDataCam
+ line_counter_config.json - line counter configuration file from OpenDataCam
+ processOpenDataCam.py - Python script file that processes the CSV files

本数据集关联的主论文为《基于微型边缘人工智能平台的实时交通量化》。
一段由DJI御Air（DJI Mavic Air）于7月28日11:30拍摄的视频，取景于南非比勒陀利亚的N4高速公路，毗邻工程4.0园区。该高清视频分辨率为4K，帧率29.97fps，时长18分45秒，文件名为2020718_Mavic_Air_4K.mp4。第二段视频为用于分析的裁剪版本（分辨率1220×560像素），文件名为20200728_Mavic_Air_Cropped.mp4。
本次分析包含两项内容：(1) 为每一辆进入画面的车辆（分为"轿车"与"卡车"两类）生成真值数据集；仅统计向西行驶（即从右向左行进）、朝向哈特菲尔德（Hatfield）的车辆。数据集包含车辆首次被观测到的帧编号及其对应类别；(2) 由OpenDataCam完成的车辆自动目标检测与计数。真值数据集（groundTruth.csv）与推理结果数据集（OpenDataCam.csv）均采用逗号分隔值（CSV）格式存储。此外还提供了OpenDataCam的行计数器配置文件（line_counter_config.json），以保障实验可复现性。同时附带了用于处理该数据集的Python脚本ProcessOpenDataCam.py。
在正确配置AI网络的阈值参数后，OpenDataCam的车辆计数准确率误差可控制在5%以内。
文件汇总如下：
+ 2020718_Mavic_Air_4K.mp4：无人机拍摄的4K分辨率原始视频文件
+ 20200728_Mavic_Air_Cropped.mp4：以高速公路为中心裁剪的4K视频文件（分辨率1220×560像素）
+ groundTruth.csv：二元车辆分类的真值数据集
+ OpenDataCam.csv：OpenDataCam输出的车辆检测与计数结果数据集
+ line_counter_config.json：OpenDataCam的行计数器配置文件
+ processOpenDataCam.py：用于处理CSV文件的Python脚本