Home

February | March | April | May | June | July | August

February

• MRes 2nd project proposal;
• Using Deep Learning to Identify Cyclists Risk Factors in London project allocation;

March

• Start of the 2nd project;
• Background reading on: cycling benefits (economy, environment, safety, equity and health perspectives), road safety indicators (accident, injury and fatality rates) and risk factors (number of intersections, vehicle speed, road width...);

April

• Background reading on: deep learning, convolution neural networks, object detection, image segmentation, UNIX programming, TensorFlow and PyTorch;

May

• Research on the available datasets containing road related detected objects/segmented images: ADE20K, MS Coco, Cityscapes and Open Images V6;
• Several object detection and image segmentation models were tested in small image datasets locally. Including YOLOv4 and YOLOv5 (s, m, l and x versions) for object detection. Xception71 and PSPNet101 for image segmentation;

June

• Based on the accuracy, performance and documentation of the previous models, YOLOv5 and PSPNet101 were selected as the object detection and image segmentation models, respectively;
• These 2 models were run in the full Google Street View imagery dataset (approximately, 1/2 million images from all London LSOAs). For YOLOv5, it was retrieved the total number of objects per image, across the 80 categories available in MS Coco Dataset. Additionaly, for PSPNet101, due to time constraints, it was only possible to obtain the total pixel counts for each of the 30 available labels in Cityscapes;
• Article published and featured in Towards Data Science about the ongoing work;

July

• Data visualization. Correlation matrices (across the multiple detected objects), histograms (providing an overview on the most common objects/labelled pixels and their distribution) and LSOA maps (representing the distribution of the MS Coco classes in Greater London) were the main tools chosen to visualize all data from the project;

August

• Article write-up. YOLOv5 real-time demo video published;
• Deadline: 20th August.