Research on automated parking perception based on a multi-sensor method

Traditional computer vision methods are used in the urban road environment, but they are vulnerable to environmental information interference and have low detection accuracy. To solve this problem, a multi-sensor method with strong anti-interference ability and desired detection accuracy is proposed in this paper. In this method, TensorFlow SSD-based deep learning is used to determine the approximate position of the parking space on a panoramic image with a marked rectangular box. In the rectangular box, grayscale, image filtering, morphological open operation, and binarization are conducted to preprocess the parking space image. Then, quadratic polynomial fitting is exploited to detect the parking lines. Next, the pixel to distance conversion formula is performed to obtain the distance coordinate of the corner relative to the vehicle center. Meanwhile, long-distance and narrow-beam angle ultrasonic radars are employed on the side of the vehicle to determine whether there is a barrier inside the parking space and detect the barrier’s position with fisheye cameras together, which helps to prevent collisions. Finally, real vehicle perception tests are conducted. The experimental results demonstrate that the proposed multi-sensor method with parallel, perpendicular, and slanted parking space recognition can resist environmental information interference effectively and achieves desired recognition accuracy.