Variable σ-AdaBoost Based Humanoid Recognition In Auto-safety System

With in the past decades, the rapid development of automobile industry and thecontinuous improvement of people’s living standard bring the car into people’s daily life.However, increasing quantity of vehicle has brought car accident to the number one killer inour moden society. Therefore, the majority of car manufactures and researchers have put morefunds in automotive active safety system that has humanoid recognition technology.Most of existing humanoid recognition systems are based on all kinds of sensors andvarieties of image processing algorithms for detecting the potential pedestrian who in front ofvehicle and the corresponding location information; with those information, systemsdetermine whether a risk of collision exists or not, furthermore, gives automatic interventionto circumvent the risk. This technique incolves plenty of interdisciplinary, such as sensors,information fusion, and machine learning, and is high in social and theoretical significance.Nowadays, many of automobile manufactories have used the latest technology in theirvehicles for protecting pedestrians. For instance, BMW Corporation puts far-infrared camerabehind cars’ grille. With the advantage of far-infrared camera of sensitiveness of objects’temperature, whole system is able to find pedestrians at distances up to300meters. Inside of100meters, the system will notice driver according to vehicle’s speed. Volvo Corporation isgoing to release a new car V40, which has been installed human-detect system andpedestrian-protect airbag. The system in V40consists of radar and optical camera. Ahumanoid detection algorithm is used to filter humanoid, and takes auto brake to avoidaccidents as speed is under35km/h. With the improvement of microelectronic imagingtechnology and lower cost of video recording equipment, the camera is destined to becomethe necessary hardware equipment of the automotive safety.Based on the collection and processing of dynamic image analysis techniques andanalysis on usage of camera of domestic and foreign automobile manufacturers. Major workscompleted in this article:1. A large number of experiments needed photos were taken by driving recoder installedin the top center of the front car’s windshield. Each of images was classified for obtaininghumanoid samples and non-humanoid samples. Many feature extraction methods were compared in the article, and Haar wavelet transform was ultimately selected for featureextraction, and60more feature points were marked for training the system.2. After analyzing the humanoid recognition methods widely, the humanoid recognitionalgorithm based on variable σ-AdaBoost SVM was introduced into this article for improvinglack of humanoid classification algorithm in other method. Embedding RBFSVM (RBF is thekernel function in SVM) into AdaBoost algorithm as weaker classifier, and using dynamicallocation samples’ weight to reduce the dependence of the classifier on the sample data,enhance the generalization ability of the classifier. As well, variable σ algorithm is trying tobalance the accuracies and the differences of each component classifiers in AdaBoost, whichimproves the classifier overall performance.3. For extraction from interesting regions, radar-assisted positioning method wasproposed in this article. With auxiliary of millimeter-wave radar, human detection system wasable to locate pedestrian region more accuracy and faster, therefore, the quantity of extractionregion was reduced, which was greatly improving the detection speed and recognitionefficiency. In order to obtain a large number of sample data, a sample collection procedurewas programmed, which could interact with training staff and elect samples with and withouthumanoid from real road pictures. All of follow-up work was done on the basis of this samplelibrary.4. Base on humanoid recognition algorithm and radar-assisted pedestrian detectionmethod above, this article built a PC-based humanoid recognition algorithm offlineexperimental verification platform. The article verified the feasibility of the proposedalgorithm and compared the performances of SVM and variable σ-AdaBoost. By simulatedradar-assisted positioning, the region of interest was located quickly, which providesprotection for applicability of the method.