A Novel Real-Time Pedestrian Detection System on Monocular Vision
Accuracy and speed are the two important keys in pedestrian detection. In order to balance these two indexes well, this thesis presents a novel pedestrian detection system, ROIs cascaded Uniform LBP and improved HOG, for real-time pedestrian detection in
- PDF / 2,303,683 Bytes
- 11 Pages / 439.37 x 666.142 pts Page_size
- 17 Downloads / 208 Views
School of Mechatronical Engineering and Automation, Shanghai University, Shanghai, China {gayshh,mhxu}@shu.edu.cn 2 Department of Electrical Engineering, Shanxi Light Industry Vocational and Technical College, Shanxi, China 3 Microelectronics Research and Development Center, Shanghai University, Shanghai, China [email protected]
Abstract. Accuracy and speed are the two important keys in pedestrian detection. In order to balance these two indexes well, this thesis presents a novel pedestrian detection system, ROIs cascaded Uniform LBP and improved HOG, for real-time pedestrian detection in monocular vision. Two contributions are made in this system. First contribution is that Uniform LBP (Local Binary Pattern) cascaded improved HOG (Histograms of Oriented Gradients) are the novel structure for pedestrian detection, which can improve detection speed. Second contribution is that this pedestrian detection system is evaluated by many methods and algorithms. Experiment shows that this system can deal with 31 fps, which can be used as the real time pedestrian detection system. Keywords: Real-time Pedestrian detection (Local Binary Pattern) LBP Improved (Histogram of Oriented Gradient) HOG (Support Vector Machine) SVM
1 Introduction Every year, the number of pedestrians who are killed in the traffic accident is incredible. These accidents cause the huge economic losses on the government and people. So, many researchers and institutions invest huge economic and humans to study the pedestrian detection in Automobile Driver Assistance System (ADAS). Pedestrian Detection (PD) gradually becomes the key technology of ADAS. The particular interest in PD has dramatically increased with the improvement of hardware technology. Although the development of PD has achieved many good results, there still doesn’t have the mature products in the market for PD with the high accuracy and speed. There are some reasons to prevent this function as the vehicle equipment. So, PD is the big problem for being embedded into the vehicle equipment. At the same time, the performance of PD is improved through more complex algorithms or the better hardware platform. Such as, Integral Channel feature [1], © Springer Science+Business Media Singapore 2016 L. Zhang et al. (Eds.): AsiaSim 2016/SCS AutumnSim 2016, Part IV, CCIS 646, pp. 293–303, 2016. DOI: 10.1007/978-981-10-2672-0_30
294
A. Guo et al.
Deformable Parts Model (DPM) [2], Conventional Neural Network (CNN) [3], Color Self Similarity (CSS) [4], Deep Learning [5], Semantic Segmentation [6] and so on [7–10] are the excellent algorithms for pedestrian detection on high accuracy. But these systems detect pedestrians too slowly on the PC. Benenson’s fast pedestrian detection system based on the GPU with CPU can process about 100 fps [11]. The hardware overload of this system is so huge, which can’t be embedded into the vehicle equipment. Others, most of the papers about pedestrian detection are realized by the OPENCV or MATLAB based on the PC [12–14]. But the drawback is that these system
Data Loading...
