An Evaluation of Boosted Features for Vehicle Detection Liwei Liu, Genquan Duan, Haizhou Ai Computer Science and Technology Department, Tsinghua University,Beijing, China ahz@mail.tsinghua.edu.cn Shihong Lao OMRON Corporation, Kyoto 619-0283, Japan lao@ari.ncl.omron.co.jp Abstract—Vehicle detection in traffic scenes is a fundamental task for intelligent transportation system and has many prac- tical applications as diverse as traffic monitoring, intelligent scheduling and autonomous navigation. In recent years, the number of detection approaches in monocular images has grown rapidly. However, most of them focus on detecting other objects (such as face, pedestrian, cat, dog, etc.) and also there lacks of vehicle datasets with various conditions for vehicle detection and comprehensive comparisons. To address these problems, we perform an extensive evaluation of many state-of-the-art detection approaches on vehicles. Our main contributions are: (1) we collect a large dataset of real-world vehicles in frontal/rear view with 30 ◦ ∼-30 ◦ yaw changes and 5 ◦ ∼ 45 ◦ pitch changes under different weather conditions (snowy, rainy, sunny and cloudy) and illumination variations, and then (2) we evaluate six types of state-of-the-art features in Real AdaBoost framework on the adequate dataset collected by ourselves and a public dataset using the same evaluation protocol. Our study presents a fair comparison and deep analysis of these features in vehicle detection. From these experiments, we explore the characteristics of good features for vehicle detection. (3) Finally, we exploit these character- istics and propose a relatively effective and efficient detector, balancing performance, speed and memory cost which can be put into practical use. I. INTRODUCTION Vehicle detection in traffic scenes is of fundamental impor- tance for surveillance system and has apparent commercial value. It exploits robust observation models for vehicle tracking which provides great potentials for many high level computer vision applications such as intelligent scheduling, traffic analysis, abnormal trajectory detection and collision avoiding. Due to its wide spread of potential applications, considerable attentions have been attracted in building auto- mated vision systems for detecting vehicles in recent decade. Many traditional works [1][2] of vehicle detection depends on background subtraction, which are sensitive to lighting variations. They also assume that vehicles appear in the scene without occlusions for a while to build up proper object models, which may encounter great challenges in crowded scenarios. In contrast to these works, machine learning based vehicle detection techniques are more robust and reliable. But they ([3][4][5]) have a short history compared with face [6][7][8][9][10] and human detection [11][12][13][14][15] and only spring up several years ago. Yet despite its ap- parent importance, the lack of comprehensive evaluations of existing popular detection techniques on vehicle detection, strongly limits the popularity in practical applications. The main reason is the lack of datasets with various conditions for vehicle detection in traffic surveillance scenes. 30° Center 0° -30° Up 5° Up 45° ... ... ... ... ... ... ... ... ... ... Up 5° Up 45° 30° -30° Horizon Fig. 1. Viewpoint range of the vehicle samples we collected. In this paper, we propose to evaluate many state-of-the- art detection approaches for vehicle on a relatively adequate dataset collected by ourselves and a public dataset. Our target is to evaluate their effectiveness and efficiency on vehicle detection and attempt to build up a detector for real world applications with well balances of speed, performance and memory cost. The most related work is [16], evaluating state- of-the-art approaches in pedestrian detection. However, not all of their performances are similar on vehicles due to the characteristics of vehicles, for example, vehicles differs much more with viewpoint changes than pedestrians as shown in Fig. 1. A good news is that since vehicles are rigid, their difference in a similar viewpoint is not huge. Therefore, the proposed evaluation on vehicles is of great significance. In the literature, there are a number of detection ap- proaches, whose key components are the training frame- work and weak features. In this paper, we do not intend to investigate various training frameworks (Boosting [17], Support Vector Machine (SVM) [11], Random Forest [18] and Latent SVM [19]), but only focus on evaluating the effectiveness of different weak features under the same training framework. Considering the speed, performance and memory issues, the selected training framework is Real AdaBoost, which has significant advantages in speed and is robust and reliable in binary classification problems as reported in [6][7][8][9][4][12]. The selected weak features can be roughly classified into grayscale based [6][10] and gradient based [11][13][14], which have achieved promising results in face and pedestrian detection respectively. Our main contribution can be summarized in three folds. (1) We collect a large dataset of vehicles from real-world for training and testing. The viewpoint is frontal/rear with