Training a classifier to detect open vehicle doors

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a classifier to detect open vehicle doors. One of the methods includes obtaining a plurality of initial training examples, each initial training example comprising (i) a sensor sample from a collection of sensor samples and (ii) data classifying the sensor sample as characterizing a vehicle that has an open door; generating a plurality of additional training examples, comprising, for each initial training example: identifying, from the collection of sensor samples, one or more additional sensor samples that were captured less than a threshold amount of time before the sensor sample in the initial training example was captured; and training the machine learning classifier on first training data that includes the initial training examples and the additional training examples to generate updated weights for the machine learning classifier.

1 . A computer-implemented method for detecting open vehicle doors, comprising:
 receiving an input sensor sample that characterizes a first vehicle and is generated from sensor data captured by one or more sensors of a second vehicle; and   processing the input sensor sample using a machine-learning classifier having a plurality of weights and having been trained in a first training process at least on first training data including a plurality of labeled training examples and a plurality of additional training examples that have been generated based on the plurality of labeled training examples to generate an open door score that represents a predicted likelihood that the first vehicle has an open door, wherein each labeled training example comprises (i) a sensor sample for which label data is available and (ii) label data classifying the sensor sample as characterizing a vehicle that has an open door, and wherein each additional training example comprises (i) an additional sensor sample that has been identified by determining that the additional sensor sample has been captured less than a threshold amount of time before the sensor sample of one of the labeled training examples and (ii) label data generated by classifying the additional sensor sample as a sensor sample that characterizes a vehicle that has an open door.   
     
     
         2 . The computer-implemented method of  claim 1 , wherein the machine-learning classifier has been further trained in a second training process on second training data including a plurality of further training examples generated using the machine-learning classifier and in accordance with updated values for the weights of the machine-learning classifier that have been updated in the first training process. 
     
     
         3 . The computer-implemented method of  claim 2 , wherein the second training process comprises using the second training data to further update weights of the machine-learning classifier starting from the updated values. 
     
     
         4 . The computer-implemented method of  claim 2 , wherein the second training process comprises using the second training data to update the weights for the machine-learning classifier starting from initial values for the weights of the machine-learning classifier. 
     
     
         5 . The computer-implemented method of  claim 2 , wherein the plurality of further training examples are generated by:
 processing each of a plurality of candidate sensor samples using the machine-learning classifier and in accordance with the updated values for the weights to generate a respective open door score for each candidate sensor sample; and   classifying each candidate sensor sample having an open door score that exceeds a threshold score as a sensor sample that characterizes a vehicle with an open door.   
     
     
         6 . The method of  claim 1 , wherein the additional sensor sample characterizes the same vehicle as the sensor sample of the one of the labeled training examples. 
     
     
         7 . The computer-implemented method of  claim 1 , wherein the sensor sample of each of the plurality of labeled training examples includes a more than a threshold amount of measurements outside of an outline of a body of the vehicle characterized by the sensor sample. 
     
     
         8 . A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations for detecting open vehicle doors, the operations comprising:
 receiving an input sensor sample that characterizes a first vehicle and is generated from sensor data captured by one or more sensors of a second vehicle; and   processing the input sensor sample using a machine-learning classifier having a plurality of weights and having been trained in a first training process at least on first training data including a plurality of labeled training examples and a plurality of additional training examples that have been generated based on the plurality of labeled training examples to generate an open door score that represents a predicted likelihood that the first vehicle has an open door, wherein each labeled training example comprises (i) a sensor sample for which label data is available and (ii) label data classifying the sensor sample as characterizing a vehicle that has an open door, and wherein each additional training example comprises (i) an additional sensor sample that has been identified by determining that the additional sensor sample has been captured less than a threshold amount of time before the sensor sample of one of the labeled training examples and (ii) label data generated by classifying the additional sensor sample as a sensor sample that characterizes a vehicle that has an open door.   
     
     
         9 . The system of  claim 8 , wherein the machine-learning classifier has been further trained in a second training process on second training data including a plurality of further training examples generated using the machine-learning classifier and in accordance with updated values for the weights of the machine-learning classifier that have been updated in the first training process. 
     
     
         10 . The system of  claim 9 , wherein the second training process comprises using the second training data to further update weights of the machine-learning classifier starting from the updated values. 
     
     
         11 . The system of  claim 9 , wherein the second training process comprises using the second training data to update the weights for the machine-learning classifier starting from initial values for the weights of the machine-learning classifier. 
     
     
         12 . The system of  claim 9 , wherein the plurality of further training examples are generated by:
 processing each of a plurality of candidate sensor samples using the machine-learning classifier and in accordance with the updated values for the weights to generate a respective open door score for each candidate sensor sample; and   classifying each candidate sensor sample having an open door score that exceeds a threshold score as a sensor sample that characterizes a vehicle with an open door.   
     
     
         13 . The system of  claim 8 , wherein the additional sensor sample characterizes the same vehicle as the sensor sample of the one of the labeled training examples. 
     
     
         14 . The system of  claim 8 , wherein the sensor sample of each of the plurality of labeled training examples includes a more than a threshold amount of measurements outside of an outline of a body of the vehicle characterized by the sensor sample. 
     
     
         15 . A non-transitory computer storage medium storing instructions that when executed by one or more computers cause the one or more computers to perform operations for detecting open vehicle doors, the operations comprising:
 receiving an input sensor sample that characterizes a first vehicle and is generated from sensor data captured by one or more sensors of a second vehicle; and   processing the input sensor sample using a machine-learning classifier having a plurality of weights and having been trained in a first training process at least on first training data including a plurality of labeled training examples and a plurality of additional training examples that have been generated based on the plurality of labeled training examples to generate an open door score that represents a predicted likelihood that the first vehicle has an open door, wherein each labeled training example comprises (i) a sensor sample for which label data is available and (ii) label data classifying the sensor sample as characterizing a vehicle that has an open door, and wherein each additional training example comprises (i) an additional sensor sample that has been identified by determining that the additional sensor sample has been captured less than a threshold amount of time before the sensor sample of one of the labeled training examples and (ii) label data generated by classifying the additional sensor sample as a sensor sample that characterizes a vehicle that has an open door.   
     
     
         16 . The non-transitory computer storage medium of  claim 15 , wherein the machine-learning classifier has been further trained in a second training process on second training data including a plurality of further training examples generated using the machine-learning classifier and in accordance with updated values for the weights of the machine-learning classifier that have been updated in the first training process. 
     
     
         17 . The non-transitory computer storage medium of  claim 16 , wherein the second training process comprises using the second training data to further update weights of the machine-learning classifier starting from the updated values. 
     
     
         18 . The non-transitory computer storage medium of  claim 16 , wherein the second training process comprises using the second training data to update the weights for the machine-learning classifier starting from initial values for the weights of the machine-learning classifier. 
     
     
         19 . The non-transitory computer storage medium of  claim 16 , wherein the plurality of further training examples are generated by:
 processing each of a plurality of candidate sensor samples using the machine-learning classifier and in accordance with the updated values for the weights to generate a respective open door score for each candidate sensor sample; and   classifying each candidate sensor sample having an open door score that exceeds a threshold score as a sensor sample that characterizes a vehicle with an open door.   
     
     
         20 . The non-transitory computer storage medium of  claim 15 , wherein the additional sensor sample characterizes the same vehicle as the sensor sample of the one of the labeled training examples.