Improving the performance of on-road vehicle detection by combining Gabor and wavelet features

Appearance-based methods represent a promising research direction to the problem of vehicle detection. These methods learn the characteristics of the vehicle class front a set of training images which capture the variability in vehicle appearance. First, training images are represented by a set of features. Then, the decision boundary between the vehicle and non-vehicle classes is computed by modelling the probability distribution of the features in each class or through learning. The purpose of this study is to Investigate the effectiveness. 8 of two important types of features for vehicle detection based on Haar wavelets and Gabor filters. In both cases, the decision boundary Is computed using Support Vector Machines (SVMs), a recent development in classification algorithms which performs structural risk minimization to maximize generalization on novel data. Both wavelet and Gabor features have demonstrated good performance in various application domains including face detection and image retrieval. Wavelet features encodo edge information, a good feature for vehicle detection. Most importantly, they capture the structure of vehicles at multiple scales. Gabor filters provide a mechanism for obtaining orientation and scale tunable edge and line detectors. Vehicles do contain strong edges and lines at different orientation and scales, thus, this type of features are also very attractive for vehicle detection. Our experimental results and comparisons using real data illustrate the effectiveness of both types of features for vehicle detection, with Gabor features performing better than Haar wavelet features. Careful examination of our results revealed that the two feature sets yield different misclassification errors which led us to the idea of combining them for improving performance. The combined set of features outperformed each feature set alone on completely novel test images, yielding an average error rate of 3.03% compared to 5.10% using Gabor features and 8.52% using Haar wavelet features.