Advanced Pedestrian Detection system using combination of Haar-like features, Adaboost algorithm and Edgelet-Shapelet

The basic task in various applications like automotive control, video surveillance, etc is human body detection. For such applications to be successful, high accuracy and high speed performance are crucial. Image feature description determines accuracy and hence it should be robust against occlusion, rotation, and changes in object shapes and illumination conditions. Till date, many such feature descriptors have been proposed. Many of them are based on histogram of oriented gradients (HOG) along with support vector machine (SVM) classifier. Limitation of this method is high time consumption though it achieved good performance for Pedestrian Detection. To counter this limitation, a Two-step framework was proposed. It consisted two steps - full-body detection (FBD) and head-shoulder detection (HSD). Zhen Li proposed fusion of Haar-like and HOG features for better performance, and HSD step utilizes Edgelet features for classification and detection. But this method results in low detection rate and less computation speed. To counter these limitations, we have proposed an advanced method to improve both detection rate and speed. We achieve this by combination of Haar-like and Triangular features for FBD and Edgelet/Shapelet for HSD. We have achieved an average 95% detection rate and 60% faster speed for this proposed method.