Embedded Vehicle Detection by Boosting

Adaptive boosting is a promising method for real time detection of vehicles for ACC applications. This paper evaluates performance and implementation issues for Adaboost classification of monocular rear view vehicle detection on embedded hardware. Images are processed on different levels, using a multi resolution band structure, and features are trained that show low evaluation complexity. Classification performance is evaluated for different types of features including orientation histograms and oriented gradient filters, with respect to receiver operating characteristics and evaluation complexity. For a selected set of negative training samples representing dense traffic scenarios, 1% false positive rate is reached at a detection rate of 95.2% using 416 operations per evaluation window