nexar-ai/nexar_collision_prediction

Nexar Collision Prediction Dataset

Dataset

The Nexar collision prediction dataset comprises videos from Nexar dashcams. Videos have a resolution of 1280x720 at 30 frames per second and typically have about 40 seconds of duration. The dataset contains 1500 videos where half show events where there was a collision or a collision was eminent (positive cases), and the other half shows regular driving (negative cases). The time of the event (collision or near-miss) is available for positive cases.

Goal

The goal of this dataset is to help build models to predict the time of collision. Models should be able to predict if a collision is about to happen or not. Both collisions and near-misses are treated equally as positive cases.

Model assessment

Models should be able to predict that a collision is about to happen as soon as possible, while avoiding false positives. Assessment scripts will be made available shortly that calculate the mean average precision across different times before collision (e.g. 500ms, 1000ms, 2000ms).

Usage

from datasets import load_dataset

dataset = load_dataset("videofolder", data_dir="/your/path/nexar_collision_prediction", split="train", drop_labels=False)

A positive example would look like this:

{'video': <decord.video_reader.VideoReader object at 0x7f127b56f940>, 'label': 1, 'time_of_event': 19.967}

and a negative example like this:

{'video': <decord.video_reader.VideoReader object at 0x7f127b56f250>, 'label': 0, 'time_of_event': None}