Our job is to determine if there is a free parking spot on parking lot or not. We got the dataset for this purpose from one parking lot in Ostrava. The photos of parking lot has been taken in diferent day time and diferent year time. On each picture there is 56 parking spots.To exctract them we use information from parking_map.txt.
We introduce 2 aproaches how the goal can be achieved.
We use sobel operator with kernel (3,3) to exctract vertical lines, then we use binary treshold. Count of non-zero pixels from result is our edge between occupied or free spot. Separately save counts for occupied and free spots. After all parking spots were processed, take lowest value from occupied spots and the highest from free spots and make an average. The result is our final edge which will be used in test phase to determine if parking spot is occupied or not.
Download OpenCV from OpenCV library and include libraries to project
Unzip the dataset.zip
Accuracy: 93.37%
We want to classify if parking spot is occupied or not, this is like made for NN classification. The base of second solution is convolutional neural network designed in Tensorflow framework with OpenCV for image pre-process. On input there is image with 1(grayscale) or 3(RGB) channels. We designed following NN:
We use softmax_cross_entropy_with_logits to get the network error and AdamOptimizer for gradient distribution with 0.001 learning rate.
Tensorflow - 1.14.0
Tensorflow GPU -1.12.0
Python OpenCV - 3.4.3.18
Numpy = 1.17.3
Unzip the dataset.zip
Accuracy: 99.25%
Training time: 6min
One parking lot evaluation time: 0.002s
One image evaluation time: 0.12s
