Object Detection And Tracking Based On Improved Hough Forest Under Traffic Condition

Object detection and tracking is one of the most popular and challenging topics in computer vision, for it is widely used in the driver assistance system and intelligent surveillance. This dissertation aims to study the vehicles and pedestrians under traffic condition. Because of the specific characteristics of vehicles and pedestrians and the complexity of actual traffic condition, there are a lot of problems in object detection and tracking, such as changes of appearance and scale, changes in perspective, changes in light intensity and partly obscured issues, etc. Now, most of the detection algorithms simplify the detection to binary classification task, that is object detection based on machine learning. Firstly train the samples by extracting their features. Secondly obtain the feature model. Thirdly search the object by sliding window in the image. In recent years, tracking-by-detection is a very active research in tracking algorithms. It converts the tracking problem to a classification problem, so that we can use a variety of discriminant models to solve it.In this dissertation, to solve vehicles and pedestrians detection and tracking under traffic condition, we build a detection and tracking system based on Hough Forest. The main work and contributions are as follows:In the feature extraction, the article uses multiple features, such as local gradient information HOG, color information Lab and texture information LBP, etc. And, improves HOG feature descriptor in order to enhance its rotational invariance and illumination invariance. The feature set extracted has a better description ability.Based on the summary of the previous theoretical research on decision tree and random forest, according to diversity of object and complexity of background, this dissertation studies and implements Hough Forest model, which combines random forest and probabilistic hough voting. This model can accommodate various discriminant information for training, so as to learn various features effectively.Innovatively, this dissertation proposes a method of updating part of random trees in the forest, makes the model capable of on-line incremental learning. Build an on-line Hough Forest model, including basic model and boosting model, furthermore, apply it to vehicles and pedestrians tracking under traffic condition.Based on off-line Hough Forest model with improved HOG and on-line Hough Forest model with incremental learning, a simulation experiment is tested on multiple data sets. Besides, the dissertation implements an object detection and object tracking system under traffic condition.