Abstract: A device performs operations including determining a probability that a vulnerable road user (VRU) will continue on a current path (e.g., in connection with controlling an autonomous vehicle). The device receives an image depicting a vulnerable road user (VRU). The device inputs at least a portion of the image into a model, and receives, as output from the model, a plurality of probabilities describing the VRU, each of the probabilities corresponding to a probability that the VRU is in a given state. The device determines, based on at least some of the plurality of probabilities, a probability that the VRU will exhibit a behavior, and outputs the probability that the VRU will exhibit the behavior to a control system.

Abstract: The systems and methods disclosed herein provide a risk prediction system that uses trained machine learning models to make predictions that a VRU will take a particular action. The system first receives, in a video stream, an image depicting a VRU operating a micro-mobility vehicle and extract the depictions from the image. The extraction process may be determined by bounding box classifiers trained to identify various VRUs and micro-mobility vehicles. The system feeds the extracted depictions to machine learning models and receives, as an output, risk profiles for the VRU and the micro-mobility vehicle. The risk profile may include data associated with the VRU/micro-mobility vehicle determined based on classifications of the VRU and the micro-mobility vehicles. The system may then generate a prediction that the VRU operating the micro-mobility vehicle will take a particular action based on the risk profile.

Abstract: Systems and methods are disclosed herein for tracking a vulnerable road user (VRU) regardless of occlusion. In an embodiment, the system captures a series of images including the VRU, and inputs each of the images into a detection model. The system receives a bounding box for each of the series of images of the VRU as output from the detection model. The system inputs each bounding box into a multi-task model, and receives as output from the multi-task model an embedding for each bounding box. The system determines, using the embeddings for each bounding box across the series of images, an indication of which of the embeddings correspond to the VRU.

Abstract: A system and a method are disclosed for determining intent of a human based on human pose. In some embodiments, a processor obtains a plurality of sequential images from a video feed, and determines respective keypoints corresponding a human in each respective image of the plurality of sequential images. The processor aggregates the respective keypoints for each respective image into a pose of the human and transmits a query to a database to find a template that matches the pose by comparing the pose to a plurality of templates poses that translate candidate poses to intent, each template corresponding to an associated intent. The processor receives a reply message from the database that either indicates an intent of the human based on a matching template, or an inability to locate the matching template, and, in response to the reply message indicating the intent of the human, outputs the intent.