Nexar Dashcam is in the business of making video recording cameras for attaching in your cars and using in other similar fields. Recently they have featured a new competition on Kaggle called the ‘Nexar Dashcam Crash Prediction Challenge’. The aim of this challenge is to create a pipeline and train a model to predict accidents 2 to 3 sec before actual impact. For training and testing purposes they have provided with 1500 mp4 video footages each; out of which half are of actual accidents and others are normal driving footages each. The final result of your model should be a probability between 0 and 1 whether the mp4 video is an actual accident or not. Some of the videos are minor impacts which need to be classified for accidents as well. The final results are calculated using 3 sets of Precision Recall Curves.
For this competition me and AM plan to work around on an approach of Feature Engineering called Optical Flow and start training model using simple LGBM gradient boosting framework. Later we can iterate or work on our approach further. This is outline for now. You can find more about our approach to the problem or concepts we’ll be using later!
