In recent years, with the booming increase of vehicle numbers in traffic, more and more accidents caused by car collision or lane departure happen everyday, which also result in fast mounting of casualties and pose a great threat to the society. Hence, more and more attention is paid to the techniques of vehicle active safety by research institudes and carmakers. Vehicle detection, vehicle collision warning, lane detection and lane departure warning techniques based on computer vision are the most important research fields of vehicle active safety systems. The vision sensor is widely used in vehicle active safety systems due to the low price, portable algorithms and the richness of the collected information. This paper focuses on vehicle active safety techniques based on monocular vision. The system processes the on-road image sequences captured by the camera mounted on mobile platform, and is able to detect the target vehicle, measure the distances between vehicles, along with lane detection. The work and contributions of the paper are as follows. 1.The paper proposes a lane detection algorithm based on current frame and the last frame. Firstly, it extracts the linear edges of the current frame with a preprocessing of making edges thicker. Then it defines a ROI region based on the position of the lane of last frame, and performs Hough transform in the ROI. After obtaining the candidates by selecting the lines which get higher votes in the mapping from the image space to the parameter space, it calculates the posterior probability based on the lane position in the last frame, and finally recognizes the lane of the current frame. 2.Studying and implementing the target vehicle detection algorithm based on GentleAdaboost. It extracts a certain number of features from the subimages, constructing a feature vector of the subregion. Then it passes the vector to the Haar cascade detector and GentleAdaboost classifier to get the result. Among them, Haar cascade detector mainly accomplishes the selection of vehicle candidates in the search area and handles the issues caused by changing in scale. The GentleAdaboost classifier classify the vehicle candidates selected by the Haar detector further and identify the frontal target vehicle with the help of lane information. 3.The paper proposes an improved method to place feature points in object tracking algorithm. While the uniform placing strategy in TLD has a serious accumulative error problem, the method we propos...
修改评论