Vehicle Counting Evaluation on Low-resolution Images using Software Tools

Vehicle counting is an important parameter in building highway macroscopic model. This model ultimately will help highway designers, road planners and even common commuters, since it can give short term predictions of the road's behaviour which is influenced for example by its traffic flow, number of lanes, as well as off and on ramps. This research attempts to count vehicles from existing video cameras, which gives low-resolution 3 seconds video of 1 frame per second. For low-resolution, conventional methods, such as Back-subtraction, and Viola Jones are unable to give high counting accuracy. However, with the aid of another custom-made software tool, various parameters of Deep Learning method, such as pixel-frame distance thresholds, and two different counting models can be run repetitively, to obtain better accuracy. Early results have shown that by varying pixel distance threshold, the percentage of error can go down from 40.8% to as low as 0.8%.