Detecting multiple objects of interest in an agricultural environment
A method includes obtaining, by the treatment system configured to implement a machine learning (ML) algorithm, one or more images of a region of an agricultural environment near the treatment system, wherein the one or more images are captured from the region of a real-world where agricultural target objects are expected to be present, determining one or more parameters for use with the ML algorithm, wherein at least one of the one or more parameters is based on one or more ML models related to identification of an agricultural object, determining a real-world target in the one or more images using the ML algorithm, wherein the ML algorithm is at least partly implemented using the one or more processors of the treatment system, and applying a treatment to the target by selectively activating the treatment mechanism based on a result of the determining the target.
1. A method implemented by a treatment system having one or more processors, a storage, and a treatment mechanism, comprising:
obtaining, by the treatment system disposed on a vehicle and configured to implement a machine learning (ML) algorithm, sensor readings of a region of an agricultural environment;
selecting a smaller portion or a layer of the sensor readings to be processed by an algorithm;
determining, by processing the smaller portion or the layer of the sensor readings, a presence of an object in the sensor readings and whether the object is a treatment candidate; and
applying, upon determining that the object is the treatment candidate, a treatment to a real-world target corresponding to the object by activating the treatment mechanism according to a treatment parameter for the treatment candidate; and
detecting a first object in a first frame at a location;
wherein, upon determining that the location of the first object has not moved out from the first frame of the sensor readings to a subsequent frame of the sensor readings, and the first object is undetected in the subsequent frame while other agricultural objects and landmarks have been detected in the subsequent frame, the first object is excluded from the treatment.
2. The method of claim 1 , wherein the smaller portion is generated by a pre-processing stage that eliminates certain tiles of sensor readings prior to the processing by the algorithm, wherein the sensor readings are divided into multiple tiles.
3. The method of claim 2 , wherein the multiple tiles comprise non-overlapping tiles.
4. The method of claim 2 , wherein the multiple tiles comprise rectangular tiles.
5. The method of claim 1 , wherein the layer of sensor readings is obtained by performing a color segmentation of the sensor readings.
6. The method of claim 1 , wherein the determining the presence of the object comprises performing multiple successive algorithmic steps, comprising applying a first ML algorithm or a first computer vision (CV) algorithm on the sensor readings followed by a second ML algorithm or a second CV algorithm on the smaller portion of the sensor readings.
7. The method of claim 6 , wherein the multiple successive algorithmic steps comprise multiple ML algorithm steps.
8. The method of claim 6 , wherein the multiple successive algorithmic steps comprise an ML algorithm step and a computer vision (CV) step.
9. The method of claim 1 , wherein an amount of treatment provided to the object is controlled based on a presence of another agricultural object within a threshold distance of the treatment candidate.
10. A treatment system having one or more processors, a storage, and a treatment mechanism, wherein the one or more processors are configured to:
obtain, by the treatment system disposed on a vehicle and configured to implement a machine learning (ML) algorithm, sensor readings of a region of an agricultural environment;
select a smaller portion or a layer of sensor readings to be processed by an algorithm;
determine, by processing the smaller portion or the layer of sensor readings, a presence of an object in the sensor readings and whether the object is a treatment candidate; and
apply, upon determining that the object is the treatment candidate, a treatment to a real-world target corresponding to the object by activating the treatment mechanism according to a treatment parameter for the treatment candidate; and
detecting a first object in a first frame at a location;
wherein, upon determining that the location of the first object has not moved out from the first frame of the sensor readings to a subsequent frame of the sensor readings, and the first object is undetected in the subsequent frame while other agricultural objects and landmarks have been detected in the subsequent frame, the first object is excluded from the treatment.
11. The treatment system of claim 10 , wherein the smaller portion is generated by a pre-processing stage that eliminates certain tiles of sensor readings prior to the processing by the algorithm, wherein the sensor readings are divided into multiple tiles.
12. The treatment system of claim 11 , wherein the multiple tiles comprise rectangular, non-overlapping tiles.
13. The treatment system of claim 10 , wherein the layer of sensor readings is obtained by performing a color segmentation of the sensor readings.
14. The treatment system of claim 10 , wherein the determining the presence of the object comprises performing multiple successive algorithmic steps, comprising applying a first ML algorithm or a first computer vision (CV) algorithm on the sensor readings followed by a second ML algorithm or a second CV algorithm on the smaller portion of the sensor readings.
15. The treatment system of claim 14 , wherein the multiple successive algorithmic steps comprise multiple ML algorithm steps.
16. The treatment system of claim 14 , wherein the multiple successive algorithmic steps comprise an ML algorithm step and a computer vision (CV) step.
17. A method implemented by a treatment system having one or more processors, a storage, and a treatment mechanism, comprising:
obtaining, by the treatment system disposed on a vehicle and configured to implement a machine learning (ML) algorithm, sensor readings of a region of an agricultural environment;
detecting, by processing the sensor readings using the ML algorithm, a first object and a second object in the sensor readings;
determining, based on features of the first object and the second object, that the first object is to be included for a treatment and the second object is to be excluded from the treatment;
selectively applying the treatment to the first object by activating the treatment mechanism in response to the determining that the first object is to be included in the treatment;
wherein the first object is detected at a location in a first frame; and
wherein, upon determining that the location of the first object has not moved out from the first frame of the sensor readings to a subsequent frame of the sensor readings, and the first object is undetected in the subsequent frame while other agricultural objects and landmarks have been detected in the subsequent frame, the first object is excluded from the treatment.
18. The method of claim 17 , wherein the sensor readings are obtained from a plurality of sensors, and wherein the sensor readings are fused together prior to the processing using the ML algorithm.
19. The method of claim 18 , wherein the plurality of sensors includes two or more of an image sensor, a sonar sensor, a radar, or a light detection and a ranging sensor (LiDAR).
20. The method of claim 17 , wherein the features comprise colors of the first object and the second object, features of the first object and the second objects, or a shade of a particular color of the first object and the second object.
21. The method of claim 17 , wherein the features comprise textures of the first object and the second object.
22. The method of claim 17 , wherein the ML algorithm is trained to identify objects to be excluded from the treatment.
23. The method of claim 17 , wherein the ML algorithm is trained to identify objects to be included in the treatment.
24. The method of claim 23 , wherein, during training of the ML algorithm, new objects that are previously unseen by the ML algorithm are added to a library and used for the training of the ML algorithm for a subsequent use.
25. The method of claim 23 , wherein the training is performed using computing resources that are offsite with respect to the agricultural environment.