Vehicle Detection Combining Objectness And Apparent Feature

Vehicle detection is an important research field of the current automotive active safety technology. The collision warning system based on vehicle detection is of great practical significance to reduce traffic accidents and improve road safety. This paper proposes a vehicle detection method which combines objectness and apparent feature. This method has characteristics of high hit rate and low false alarm rate, and performs efficiently with single frame images.First, we study the on-road camera's imaging principle, and make a conclusion that the vehicles in a image are appearing along the vanishing line of the road, so that we can focus the region of interest in a image on the neighbourhood of the vanishing line and create many detection windows at different scales along the vanishing line, which reduces the search space significantly.Next, objectness is introduced as a part of hypothesis generation of vehicle detection, and we select salience, color, edges and super-pixels as the cues of objectness, and give the scoring rule of each cue. We combine the four clues and score every detection window in the region of interest. Then we select several windows with high probability of covering an object according to their scores, and fuse them using the method of non-maxima suppression. We locate the hypothesized vehicles in the image with new windows, which are considered as the output of hypothesis generation stage.Finally, we achieve hypothesis validation by extracting HOG features from the output windows of hypothesis generation, and classify them with trained multi-scale classifiers. In the hypothesis validation stage we focus on HOG feature extraction algorithm and multi-scale training. After hypothesis validation we propose a method based on edge information to estimate the distance from the front vehicles and ego position.The proposed method is tested in various scenes, and achieved very good performance, which meet the real-time applications nicely.