System and method for railway weeding

A weeding system for a railway weeding vehicle, comprising:

• at least one camera adapted to be mounted on a railway vehicle to acquire an image of a portion of a railway track while said railway vehicle is travelling on a train track, said image comprising a matrix of pixel values,
• a spraying unit adapted to be mounted on said railway vehicle,
• a control unit adapted to receive images from said camera, generate a weed species detection signal and selectively cause the spraying of a chemical agent by the spraying unit on the basis of said weed species detection signal,
• the control unit being adapted to execute in a running mode a training-based algorithm, the training being based on a set of reference images labeled so as to indicate the effective presence or not of at least one weed species in said images,
• wherein each of said reference images comprises a pair of a nighttime image and a daytime image of the same scene, said labeling is performed on the daytime images and applied to the corresponding nighttime images, the training is based on the labeled nighttime images, and in the running mode, the algorithm receives nighttime real images,
• wherein each pair of a reference daytime image and nighttime image comprises a real nighttime image and a simulated daytime image generated from said nighttime image by means of a training-based conversion algorithm, and
• wherein said conversion algorithm is executed on raw Bayer images from a digital camera and trained by pairs of real day time and nighttime images of the same scene in raw Bayer mode.

A method for training a training-based weed recognition algorithm for use in a railway weeding vehicle, the railway vehicle comprising:

• at least one camera adapted to acquire an image of a portion of a railway track while said railway vehicle is travelling on a train track, said image comprising a matrix of pixel values,
• a spraying unit,
• a control unit adapted to receive running-mode nighttime images from said camera, generate a weed species detection signal and selectively cause the spraying of a chemical agent by the spraying unit on the basis of said weed species detection signal,
• the method comprising training the weed recognition algorithm based on pairs of a daytime image and a nighttime image of the same scene, the training comprising the following steps:
• determining a label from each daytime image of a pair based on the presence or not of weed species in said daytime image,
• assigning said label of the corresponding nighttime image of the same pair,
• using the labeled nighttime images for training the weed recognition algorithm, and
• determining a weed species contour from each daytime image of a pair, said contour being applied to the corresponding nighttime image.

A method for training a training-based weed recognition algorithm for use in a railway weeding vehicle, the railway vehicle comprising:

• at least one camera adapted to acquire an image of a portion of a railway track while said railway vehicle is travelling on a train track, said image comprising a matrix of pixel values,
• a spraying unit,
• a control unit adapted to receive running-mode nighttime images from said camera, generate a weed species detection signal and selectively cause the spraying of a chemical agent by the spraying unit on the basis of said weed species detection signal,
• the method comprising training the weed recognition algorithm based on pairs of a so-called daytime image and a so-called nighttime image of the same scene, the training comprising the following steps:
• determining a label from each daytime image of a pair based on the presence or not of weed species in said daytime image,
• assigning said label of the corresponding nighttime image of the same pair,
• using the labeled nighttime images for training the weed recognition algorithm,
• wherein each pair of a reference daytime image and nighttime image comprises a real nighttime image and a simulated daytime image generated from said nighttime image by means of a training-based conversion algorithm, and
• wherein said conversion algorithm is executed on raw Bayer images from a digital camera and is trained by pairs of real day time and nighttime images of the same scene in raw Bayer mode.