Presumably if you are making decisions on probabilities, you meant that you wanted to use threshold of of 0.5 and not a threshold of 50? 

If the SVM is predicting probabilities, thresholding is very simple. If you are using Matlab, you can make binary predictions with one line: decisions = probs > 0.5;. Other languages should be similar. 

What is EER? If you want to performance metrics, you should consider computing the following: 

• Accuracy;
• Precision;
• Recall;
• F1 score (especially if your class distribution is unbalanced); and 
• Brier score

A threshold of 0.5 is almost certain sub-optimal, however. If the classes are unbalanced, or if misclassifying one class is more costly than misclassifying the other, it is worth looking into the field of cost-sensitive learning to select optimal thresholds. I'm attaching a seminal paper on the topic ("The foundations of cost-sensitive learning".).

Also, LibSVM will employ Platt scaling for probability estimation, but it is well known that SVM probability estimates are not well calibrated. It might also be worth looking at other probability calibration methods (e.g. "Transforming classifier scores into accurate multiclass probability estimates.").

Thanks for your interest in the question. I did mean threshold of 0.5. The data is balanced ,hence 0.5 threshold will be optimal. I am using LIBSVM in MATLAB. Please check my understanding-after i train the model using svmtrain command, i use trained model to predict labels for test data using command

[~,~,p]=svmpredict(testlable,testdata,model, '-b 1'. )

Then i should assign labels to test data 1 if p >0.5 and 0 if p

Equal Error Rate is the point where false positive rate is equal to false negative rate (hence equal errors), i.e. FPR = FNR = 1 - TPR.

I am not familiar with matlab, but as you already have an ROC curve plotted (I assume the data is available then), why not iterating over the data and finding the ones satisfying the equation (or possibly the points around it)?