Research&Implementation Of Backover Obstacle Detection Based On Monocular Vision

With the popularity of the car in recent years, the problems due to traffic accidents have become prominent, while the accidents caused by parking have a large proportion. Therefore, it is essential to identify the surrounding environment’s useful information accurately. Most of the existing parking assist products are based on radar or ultrasonic sensors, it has several weak points, such as small detection area, blind spots, having a bad performance in detecting pedestrians, especially children and other small objects, so obstacle detection system based on vision has become a hot point in the parking assist research area.After analyzing and comparing lots of obstacle detection algorithms, the paper has used different algorithm according to the vehicle’s state. When the vehicle is still, a background subtraction approach is used for obstacle detection. And if the background model is not fully finished, a frame difference approach is used. When the vehicle is moving, the paper has proposed an obstacle detection algorithm based on multi-frame. The algorithm firstly build ego vehicle motion model, search for current frame projection point amongst the latest multi-frame, using the ego-vehicle motion parameters which are obtained from motion sensor. Secondly, building the distribution model for each current frame pixel using corresponding points. Thirdly, taking advantage of the difference between obstacle and road plane motion to detect the current frame pixel belong to road plane (or obstacles), generating the binary result image. Finally, withdrawing the obstacle region through eliminating noise points in binary result image.The proposed algorithm in the paper has been implemented in Recognition Engine which belongs to Neusoft Advanced Automotive Electronic Research Center. The algorithm has been tested in real world environments and has achieved expected results in recognition rate and performance.