Sensor fusion for collision detection

1. A system for collision detection, comprising:

an interface configured to receive sensor data from a plurality of sensors; and

a processor configured to:

preprocess the sensor data using a set of preprocessing steps to create preprocessed sensor data, wherein the set of preprocessing steps comprises filtering, normalization, and alignment;

process the preprocessed sensor data using a compound model to create a collision score, wherein a combined input generated from a first neural network output and a second neural network output are processed using a dense layer neural network of the compound model to generate the collision score; and

provide the collision score.

2. The system of claim 1 , wherein the plurality of sensors comprises an accelerometer, a gyroscope, a GPS, and an audio sensor.

3. The system of claim 1 , wherein the processor is further configured to determine a collision indication indicating whether a collision occurred based at least in part on the collision score.

4. The system of claim 3 , wherein the processor is further configured to provide the collision indication.

5. The system of claim 1 , wherein the processor is further configured to determine and/or provide a collision severity.

6. The system of claim 1 , wherein the compound model comprises a neural network model, a machine learning model, a deep learning model, and/or a plurality of models.

7. The system of claim 1 , wherein the processor is further configured to provide an indication to take an action based at least in part on the collision score, wherein the action comprises one or more of: fetching data, providing data to a human reviewer, providing a notification to a client, storing data, creating a report, initiating a 911 call, initiating driver contact for a client operator, providing a tow truck request, providing a backup truck for cargo transfer request, or providing an ambulance request.

8. The system of claim 1 , wherein alignment comprises timestamp alignment or peak alignment.

9. The system of claim 8 , wherein timestamp alignment comprises data interpolation.

10. The system of claim 8 , wherein peak alignment comprises positioning of a data peak at a predetermined position within a data sample.

11. The system of claim 1 , wherein the set of preprocessing steps additionally comprises bias removal, subsampling, sample shortening, sample padding, sample truncating, spectrogramming, and/or mel spectrogramming.

12. The system of claim 1 , wherein sensor data comprises audio data that is preprocessed using filtering, length correction, alignment and/or mel spectrogramming.

13. The system of claim 1 , wherein a first set of sensors preprocessed data are input to the first neural network of the compound model to generate a first neural network output.

14. The system of claim 1 , wherein a second set of sensors preprocessed data is input to the second neural network of the compound model to generate a second neural network output.

15. The system of claim 1 , wherein the model is trained using human reviewer data.

16. The system of claim 1 , wherein the model is trained using binary data indicating whether or not a crash occurred.

17. A method for collision detection, comprising:

receiving sensor data from a plurality of sensors;

preprocessing, using a processor, the sensor data using a set of preprocessing steps to create preprocessed sensor data, wherein the set of preprocessing steps comprises filtering, normalization, and alignment;

processing the preprocessed sensor data using a compound model to create a collision score, wherein a combined input generated from a first neural network output and a second neural network output are processed using a dense layer neural network of the compound model to generate the collision score; and

providing the collision score.

18. A computer program product for collision detection, the computer program product being embodied in a non-transitory computer readable storage medium and comprising computer instructions for:

receiving sensor data from a plurality of sensors;

preprocessing the sensor data using a set of preprocessing steps to create preprocessed sensor data, wherein the set of preprocessing steps comprises filtering, normalization, and alignment;

processing the preprocessed sensor data using a compound model to create a collision score, wherein a combined input generated from a first neural network output and a second neural network output are processed using a dense layer neural network of the compound model to generate the collision score; and

providing the collision score.