Carrying more passengers than capacity limited to the bus is still a problem in Ethiopia. Passengers are obliged to be squeezed in the spaces between the seats and vehicle seats are expected to carry people beyond their limit. This bribery act is accomplished by the drivers which is becoming against the laws outlined by the transport authority of the country. Despite having some stationing police control at different locations for manual checkup, the increase in prevalence of excessive passengers still persists. This brings the task of automatic passenger flow estimation very useful in public transportation system. In this thesis work a combined convolutional neural network-based passenger detection model and a computer vision-based tracking algorithm are developed for an automatic passenger counting system to alleviate the problems stated. SSD detection model is designed for the detection task and a Dlib tracker typically for avoiding double counting of passengers is established. A PRI camera module is used for providing the images in the subsequent frames to the RPI processor. Alternatively, to decrease the execution time of the processor, video file containing group of people entering and leaving an area will be given to the processor. To decrease the operation time of the processor, a video file with six people is taken and a threshold of four is used for the throttle valve to take action. When the number of people counted by the designed machine learning based algorithm reaches the carrying capacity of the vehicle; that is the threshold value, the throttle valve of the engine will be closed for blocking the air flow which intern decrease the power of the engine. An engine with minimum power will not have a capacity to move the vehicle and consequently the vehicle will stop. The prototype is tested on an academic purpose electrically operated engine. The engine will not stop since the engine is not fueled engine. Therefore, a separate control algorithm is provided for the engine power input when the throttle valve change position from fully open to fully closed. The system has an accuracy of 91 % percent when the images are clear and large enough to be detected.