Methods and systems to predict activity in a sequence of images

1. A method to determine activity in a sequence of successively acquired images of a scene, comprising:

acquiring the sequence of images;

forming, for each image in the sequence of images, a feature block of features extracted from the image;

determining, for each image in the sequence of images, image specific information, wherein the image specific information includes a weighting indicating Image importance for the image and one or more likelihoods of one or more activities;

normalizing, for each image in the sequence of images, the weighting to form a normalized weighting;

combining, for each image in the sequence of images, the normalized weighting and feature block to form a weighted feature block;

passing a plurality of weighted feature blocks through a predictive module to determine an activity in the sequence of images;

comparing the determined activities with a most likely image activity in the image specific information;

validating the determined activity based on the comparison; and

outputting a result comprising the validated activity in the sequence of images.

2. The method of claim 1 , wherein determining image specific information comprises:

determining respective likelihoods of a plurality of predetermined activities occurring in the image, and the weighting for the image relates to the highest of the determined likelihoods of the plurality of predetermined activities.

3. A method to determine activity in a sequence of successively acquired images of a scene, comprising:

acquiring the sequence of images;

forming, for each image in the sequence of images, a feature block of features extracted from the image;

determining, for each image in the sequence of images, image specific information, wherein the image specific information includes a weighting indicating Image importance for the image and one or more likelihoods of one or more activities;

storing the formed feature blocks and the determined image specific information, wherein forming the feature block of features extracted from the image and determining the image specific information comprises retrieving a previously determined feature block and image specific information for the image;

normalizing, for each image in the sequence of images, the weighting to form a normalized weighting;

combining, for each image in the sequence of images, the normalized weighting and feature block to form a weighted feature block;

passing a plurality of weighted feature blocks through a predictive module to determine an activity in the sequence of images; and

outputting a result comprising the determined activity in the sequence of images.

4. The method of claim 1 , further comprises:

triggering a further action based on the comparison.

5. The method of claim 4 , wherein triggering the further action comprises at least one of:

issuing a warning; and

adjusting the determined activity to a warning or a default value.

6. The method of claim 4 , wherein triggering the further action comprises:

determining a number of disagreeing images based on identifying a set of compared images that reveal a difference in the comparison; and

responsive to the number of disagreeing images being greater than half of the number of images in the sequence, outputting a updated activity in the sequence of images.

7. The method of claim 6 , wherein outputting the updated activity in the sequence of images comprises:

determining, based on the image specification specific information, the most likely image activity for each of the set of compared images is the same; and

outputting the most likely image activity as the activity in the sequence of images.

8. The method of claim 2 , wherein the result comprises the image specific information for at least one frame.

9. The method of claim 1 , wherein forming the feature block of features extracted from the image and determining the image specific information comprises:

passing each image in the sequence of images through a feature encoding convolutional neural network to form a plurality of feature blocks; and

passing each feature block of the plurality of feature blocks through an image-based module comprising at least one fully connected layer.

10. The method of claim 1 , wherein passing the plurality of weighted feature blocks through the predictive module comprises:

passing a concatenation of the weighted feature blocks through a time-based module comprising a convolutional neural network.

11. The method of claim 1 , wherein normalizing the weighting to form the normalized weighting comprises passing the weighting through a SoftMax module.

12. A method for training a predictive module to determine activity in a sequence of successively acquired images of a scene, the method comprising:

a) acquiring a sequence of images of a scene labelled according to an activity in the scene;

b) processing the sequence including forming a feature block of features extracted from each image and determining image specific information including a weighting for the image and an activity prediction for the image;

c) normalizing the determined weightings to form a normalized weighting for each image in the sequence of images;

d) for each image in the sequence of images, combining the associated normalized weighting and associated feature block to form a weighted feature block;

e) passing a combination of the weighted feature blocks through a predictive module to determine an activity in the sequence of images;

f) for each image in the sequence, comparing the activity prediction for the image to the label for the image, and using the result of the comparison to improve the activity prediction for the image;

g) comparing the determined activity in the sequence of images to the label for the sequence and using the result of the comparison to improve the determination of the activity in the sequence of images; and

h) repeating steps a) to g) for successive sequences of images until a condition is met.

13. The method of claim 12 wherein the condition comprises either: each prediction being within a threshold value of a label value; or a set number of repetitions of steps a) to g) has been performed.

14. A camera system, comprising:

sensors;

memory; and

one or more processors configured to perform operations comprising:

acquiring a sequence of images;

forming, for individual images in the sequence of images, a feature block of features extracted from the image;

determining, for individual images in the sequence of images, image specific information, wherein the image specific information includes a weighting for the image and one or more likelihoods of one or more activities;

normalizing, for individual images in the sequence of images, the weighting to form a normalized weighting;

combining, for individual images in the sequence of images, the normalized weighting and feature block to form a weighted feature block;

passing a plurality of weighted feature blocks through a predictive module to determine an activity in the sequence of images;

comparing the determined activities with a most likely image activity in the image specific information;

validating the determined activity based on the comparison; and

outputting a result comprising the validated activity in the sequence of images.

15. The camera system of claim 14 , further comprises a driver-facing camera configured to generate the sequence of images, wherein the predictive module configured to receive the sequence of images.

16. The camera system of claim 14 , the operations further comprising:

causing a vehicle to modify a vehicular system according to the result.

17. The camera system of claim 14 , wherein the predictive module is further configured to determine likelihoods of predetermined activities occurring in individual images in the sequence of images, and wherein the result comprises the determined activity in a sequence of images and the image specific information for at least one frame.