Realtime-Action-Recognition

Realtime-Action-Recognition copied to clipboard

@li0128 Hi, 276 is too long to draw on the image ... So I just draw the last digit of it on the image, which is 6. By the way, thank you for testing my program on the real-world scenarios. I guess it's pretty bad, right? I tested it once, and the result makes me so ashamed (T＿T)

@li0128 About the P4 and P5 in your figure, they don't have labels because I added a criterion to remove people who are too small in the image. You may try and modify it. It's in def remove_skeletons_with_few_joints(skeletons) in src/s5_test.py.

        if num_valid_joints >= 5 and total_size >= 0.1 and num_leg_joints >= 0:
            # add this skeleton only when all requirements are satisfied
            good_skeletons.append(skeleton)

@felixchenfy Thank you very much ! The paper "ST-GCN" can learn more features automatically to classify actions .

I tried this on a surveillance cam, the id of people get swapped when two people cross path. Also if in frame t 2 person is detected and frame t+1 only 1 person is detected both the person share same ID.

NB : I have made few changes to implement deepsort in this, I am not much of a expert in computer vision, so I might have done something wrong. If someone is interested and want to look into my codes, I am much happy to share it with you.

I tried this on a surveillance cam, the id of people get swapped when two people cross path. Also if in frame t 2 person is detected and frame t+1 only 1 person is detected both the person share same ID.

NB : I have made few changes to implement deepsort in this, I am not much of a expert in computer vision, so I might have done something wrong. If someone is interested and want to look into my codes, I am much happy to share it with you.

Hi althulvingt, I am working follow this tutorial too, could you please share your codes with changing in deepsort tracking? Thank you in advance,