Color image vehicular detection systems with and without fuzzy neural network: A comparison

This paper develops a color image vehicular detection (CIVD) system in which background differencing technique is employed to detect whether a vehicle passes through the detecting points equally spaced out on a pseudo line detector. Two methods (interval search and regression) are tried to determine the optimal crisp threshold values to cope with various lighting conditions. To compare the detection performance with and without incorporating a fuzzy neural network (FNN), a three-layer FNNCIVD system is further designed with trapezoidal membership function and network parameters trained by back propagation algorithm. Under different environments (freeway and urban street) with various lighting conditions (daytime and nighttime), it is found that the detection success rates for interval-search CIVD and regression CIVD are about the same. However, both perform worse than the FNNCIVD system in which about 90% success rates are reported with seven detection points. Compared with the interval-search CIVD system, the FNNCIVD system can increase the success rates at a range of 14% to 22% on the freeway mainline and 18% to 26% on the urban street. It is also found that daytime detection performance is slightly better than nighttime detection. Possible reasons for missed detection and false detection are discussed.