The invention, which belongs to the field of machine vision technology, discloses a real-time detection method for determining occupancy of a parking space and particularly relates to a real-time detection method for determining whether a parking space is occupied correctly based on fusion of an overlooking deep image with a convolution neural network. A binocular stereo camera is built; a vehiclesample and a non-vehicle sample are collected; a vehicle detection model based on a convolutional neural network CarNet is designed and the vehicle detection model is trained; the location of a monitored parking space in a camera image is calibrated manually; a deep image in a monitoring scene is obtained; the deep image is segmented to obtain a vehicle area preliminarily and the vehicle is detected by using the CarNet detection model; and on the basis of combination of the deep image with the convoluted neural network result, a reliable vehicle detection frame is obtained and whether the parking space is occupied is determined. The real-time detection method can be applied to semi-open or indoor parking lots like a gas station or an underground parking lot and is capable of outputting parking space occupation information accurately. The method has advantages of low implementation cost, high work efficiency and strong automation capability and has broad development prospects.