Reliable vehicle type classification by Classified Vector Quantization

A working vehicle detection and classification system is proposed in this paper. The vehicle detection is implemented by a multiple layer perceptrons (MLP) ensemble using Haar-like features. To address the classification reliability issue, a prototype based scheme, called Classified Vector Quantization (CVQ), was applied for vehicle classification. By CVQ, each data category is represented by its own codebook, which can be implemented by some efficient neural learning algorithms, for example, the self-organizing map (SOM) and neural 'gas' algorithm. In classification process, each codebook offers a generalized 'nearest neighbor' by a population decoding principle to be compared with the input data. The advantage of CVQ is its convenience to provide reliable classification using the embedded rejection option. Experiments demonstrated the efficiency for vehicle classification task. The scheme offers a performance of accuracy over 95% with a rejection rate 8% and reliability over 98% with a rejection rate 20%. This exhibits promising potentials for real-world applications.