SYSTEMS AND METHODS FOR DETECTING INTERSECTION CROSSING EVENTS USING FULL FRAME CLASSIFICATION TECHNIQUES
Disclosed herein are systems and methods for classifying traffic indicators to detect intersection crossings by a vehicle. A computing device can receive a sequence of frames captured by a capture device mounted to the vehicle, and generate an intersection status data structure for each frame in the sequence of frames using a full-frame classification model. The computing device can also classify a set of features detected in each frame using an object detection model, and detect an intersection crossing event based on the intersection status data structure of each frame and the classification of each of the set of features detected in each frame. The systems and methods described herein improve upon existing traffic light classification models by using a full-frame intersection classification model in connection with traditional object tracking techniques, thereby reducing false-positive crossing-event detection and improving the accuracy of classification in a variety of environmental conditions.
1 . A method to detect intersection crossings by a vehicle, the method comprising:
receiving, by one or more processors coupled to a memory, a sequence of frames captured by an image capture device mounted to the vehicle;
generating, by the one or more processors, an intersection status data structure for each frame in the sequence of frames using a classification model trained on a labeled dataset of frames depicting roadways, wherein each label of the labeled dataset used to train the classification model is applied to a substantial portion of an image frame;
classifying, by the one or more processors, using an object detection model, a set of features detected in each frame of the sequence of frames; and
detecting, by the one or more processors, an intersection crossing event based on the intersection status data structure of each frame and the classification of each of the set of features detected in each frame.
2 . The method according to claim 1 , wherein the object detection model comprises a neural network that comprises at least one layer trained for detecting an object in the image frame.
3 . The method according to claim 1 , wherein the classification model comprises a neural network that comprises at least one layer trained for detecting a traffic light in the image frame.
4 . The method according to claim 1 , wherein the classification model comprises a plurality of output heads corresponding to characteristics of the intersection in the image frame.
5 . The method according to claim 4 , wherein at least two of the plurality of output heads are trained on a common set of training data.
6 . The method according to claim 4 , wherein the plurality of output heads comprises at least one of an intersection type, a crossing type, a first status of the traffic light for a left turn, a second status of the traffic light for a right turn, or a third status of the traffic light for going straight through the intersection.
7 . The method according to claim 1 , wherein the one or more processors execute logic combining a classification model output of an intersection crossing event based on the intersection status data structure of each frame and an object detection model output of the classification of each of the set of features detected in each frame.
8 . The method according to claim 7 , further comprising inputting, by the one or more processors, the sequence of frames into an object detection layer and a traffic light detection layer at a same time.
9 . The method according to claim 1 , wherein each label comprises an indication of whether the vehicle is approaching an intersection, within an intersection or not at an intersection.
10 . A system for detecting intersection crossings by a vehicle, the system comprising one or more processors coupled to a non-transitory memory, the one or more processors configured to perform the steps of:
receiving a sequence of frames captured by an image capture device mounted to the vehicle;
generating an intersection status data structure for each frame in the sequence of frames using a classification model trained on a labeled dataset of frames depicting roadways, wherein each label of the labeled dataset used to train the classification model is applied to a substantial portion of an image frame;
classifying using an object detection model, a set of features detected in each frame of the sequence of frames; and
detecting an intersection crossing event based on the intersection status data structure of each frame and the classification of each of the set of features detected in each frame.