In some implementations, a device may receive video data associated with video frames that depict an environment of a vehicle. The device may identify an object depicted in the video frames, wherein an object detection model indicates bounding boxes associated with the object. The device may determine, based on the bounding boxes, a configuration of the bounding boxes within the video frames. The device may determine, based on the configuration of the bounding boxes, that the object is a vulnerable road user (VRU) that is in the environment. The device may determine a trajectory of the VRU based on a change in the configuration between video frames of a set of the video frames. The device may determine, based on the trajectory, a probability of a collision between the VRU and the vehicle. The device may perform, based on the probability, an action associated with the vehicle.