A vision-based vehicle classifier which uses machine learning techniques to identify cars and trucks in video images.
Continental Automotive built a driver assistance system to help drivers safely change lanes, merge, avoid obstacles, and perform similar maneuvers. NREC’s vehicle classifier was part of Continental’s future active safety systems. The camera from this system scans the road ahead, looking for other vehicles. NREC's classifier examines these images and uses highly efficient machine learning techniques to quickly and efficiently find areas that contain vehicles.
Intelligent assistance systems that sense a car or truck’s surrounding and provide feedback to the driver can help to prevent accidents before they happen and will make driving safer, easier, and less tiring.
Current driver assistance systems use vision, ladar, or radar to perceive the vehicle’s environment. However, these sensors all have drawbacks. Radar-based systems detect the proximity of other vehicles but do not give a detailed picture of the surroundings. Ladar-based systems also detect proximity but may not work well in bad weather. Vision-based systems provide detailed information, but processing and interpreting images in the short time frame needed to make driving decisions is challenging.
Continental Automotive Systems developed vision-based driver assistance systems to help drivers avoid accidents. Continental drew upon NREC’s expertise in machine vision and machine learning to develop a classifier that quickly and efficiently detects the presence of other vehicles on the road. NREC’s vehicle classifier is designed to identify and locate vehicles in real-time video images from the lane departure warning system’s camera.
The vehicle classifier uses a fast, computationally efficient classification algorithm to identify which images contain vehicles and which ones do not. It runs on an inexpensive digital signal processor (DSP). Its input is video from the lane departure warning system’s camera, located behind the driver’s rearview mirror.
The classification algorithm trained on a data set that contains video images of roads with cars, trucks and other vehicles. The locations of each vehicle are labeled by hand in every video frame of the training data set. From this data set, the algorithm learns which image features represent other vehicles and which ones do not. NREC developed algorithms that reduced the training time by over an order of magnitude from previously published results.
The classification algorithm scans raw incoming images to identify regions in the image that contain cars. An important characteristic of this algorithm is that it works extremely quickly, allowing large regions of the image to be processed at the video frame rate.